Poverty & Equity Global Practice Working Paper 212 WHERE THEY LIVE DISTRICT-LEVEL MEASURES OF POVERTY, AVERAGE CONSUMPTION, AND THE MIDDLE CLASS IN CENTRAL ASIA William Seitz July 2019 Poverty & Equity Global Practice Working Paper 212 ABSTRACT Rapid economic growth over the past two decades lifted millions of people out of poverty in Central Asia. But the uneven spread of prosperity left many communities struggling to catch up. To support lagging regions within countries, each of the region’s five national governments has made convergence a pillar of their development strategies. An imperfect patchwork of household surveys allows policy makers to monitor progress and identify some spatial disparities. But these share an important weakness: none of the official surveys in the region is representative when disaggregated to the level of districts. Islands of poverty and prosperity are thus lost in the averages—leading to targeting inaccuracies that can slow the pace of poverty reduction. This study partially addresses the challenge. The accuracy of key welfare indicators is sharpened well beyond what could be achieved for any country alone by: i) unifying survey data from across the region and ii) applying the techniques of small-area estimation. The results provide detailed measures of welfare that in turn can be disaggregated for each district in Central Asia. Comprehensive maps of where the poor and the middle class live are presented, for the entire region and individually for each country. This paper is a product of the Poverty and Equity Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and contribute to development policy discussions around the world. The author may be contacted at wseitz@worldbank.org. The Poverty & Equity Global Practice 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. ‒ Poverty & Equity Global Practice Knowledge Management & Learning Team This paper is co-published with the World Bank Policy Research Working Papers. Where They Live District-Level Measures of Poverty, Average Consumption, and the Middle Class in Central Asia William Seitz1 Keywords: Poverty mapping. Fay-Herriot, Small Area Estimation, Central Asia, Regional Development JEL Codes: I30, I32 R12 1 Poverty and Equity Global Practice; the World Bank; wseitz@worldbank.org ACKNOWLEDGMENTS This report is a product of the World Bank Poverty and Equity Program for Central Asia, led by William Seitz (Economist, GPV03). Maps and data files were generated with assistance from Avralt- Od Purevjav (Consultant, GPV03). The approach used to derive the small area estimates in this report relies on a Stata command developed by Paul Andres Corral Rodas (Data Scientist, GPV03), William Seitz, Minh Cong Nguyen (Economist, GPV03) and Joao Pedro Azevedo (Lead Economist). Reviewers include: Bob Fay (Senior Scientist, Westat), Alexandru Cojocaru (Senior Economist, GPV03), Kenneth Simler (Senior Economist, GPV02), Kristen Himelein (Senior Economist, GPV02) and David Newhouse (Senior Economist, GPV07). Additional project support was provided by Alisher Rajabov (Economist, GPV03), and Saida Ismailakhunova (Senior Economist, GPV03). This task was completed under the guidance of Carlos Silva-Jauregui (Lead Economist and Acting Practice Manager, GPV03) 2  Contents I – Introduction ............................................................................................................................................. 7  I.I – Definitions of Poverty in Central Asia ........................................................................................... 8  I.II – The Middle Class in Kazakhstan, and Average Per Capita Consumption .............................. 15  II – Data ...................................................................................................................................................... 17  II.I – Districts by Country  ..................................................................................................................... 18  II.II – Kazakhstan HBS data for 2017 .................................................................................................. 18  II.III – Uzbekistan L2CU Baseline Survey 2018 ................................................................................. 19  II.IV – Kyrgyz Republic HBS 2016  ....................................................................................................... 19  II.V – Tajikistan L2T Baseline Survey for 2015 ................................................................................... 20  II.VI – Satellite and Administrative Data ............................................................................................ 20  III – Fay-Herriot Small Area Estimation Model ........................................................................................... 21  IV – Results ................................................................................................................................................. 23  V – Concluding Remarks and Example Applications  .................................................................................... 28  VI – References ........................................................................................................................................... 32  VII – Annexes  .............................................................................................................................................. 33  Annex A – Kazakhstan District-level Results 2017 .............................................................................. 33  Annex B – Uzbekistan District-level Results 2018 .............................................................................. 40  Annex C – Kyrgyz Republic District-level Results 2016 ..................................................................... 46  Annex D – Tajikistan District-level Results 2015 ................................................................................ 48  Annex E – Country-Level Maps ........................................................................................................... 51  Annex F – Validation and Precision ..................................................................................................... 66  Annex G – Comparison of Domain Variance Estimation Alternatives ............................................. 72  Annex H – Regression Model Details  .................................................................................................. 74  Annex I – Sensitivity to Model Selection Technique .......................................................................... 76  Annex J – K-fold Model Selection Sensitivity Analysis ....................................................................... 78  3    Tables Table 1: Survey Data Used ........................................................................................................................ 18  Table 2: Kazakhstan District-level Results 2017 ...................................................................................... 33  Table 3:Uzbekistan District-level Results 2018 ....................................................................................... 40  Table 4: Kyrgyz Republic District-level Results 2016 ............................................................................. 46  Table 5:Tajikistan District-level Results 2015 ......................................................................................... 48  Table 6: Fay-Herriot Model for Poverty Rate at $5.5/day and $3.2/day ............................................... 74  Table 7: Fay-Herriot Model for share below middle-class line, and average consumption in 2011 PPP  .................................................................................................................................................................... 75  4    Figures Figure 1: Cross-Country Comparison of Income Class Poverty Rates .................................................. 10  Figure 2: Poverty rates in Kazakhstan using international poverty lines .............................................. 11  Figure 3: National Poverty Rates in Tajikistan (left-note not comparable to international poverty estimates); Number of quarters in poverty (right) .................................................................................. 12  Figure 4: Poverty rates in the Kyrgyz Republic ....................................................................................... 13  Figure 5: Poverty Rates by Region of Uzbekistan (2018) ....................................................................... 14  Figure 6: UMIC Poverty Rate by Regions (Oblasts) in Central Asia .................................................... 15  Figure 7: The middle class share of the population, by regions (oblasts) of Central Asia ................... 16  Figure 8: Average Daily Per Capita Consumption in 2011 PPP ............................................................. 17  Figure 9: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Central Asia ..................... 24  Figure 10: Poverty at $5.5-a-day PPP per person, rates for Central Asia ............................................... 25  Figure 11: Average Per Capita Consumption 2011 PPP .......................................................................... 26  Figure 12: Share of Population in the Middle Class in Central Asia ...................................................... 27  Figure 13: Share of Households with at Least One Member Abroad for Work .................................... 28  Figure 14: District Share of HHs with Migrants by District Per Capita Consumption Quintile  ......... 29  Figure 15: Number of Early Learning Centers per 1000 Children  .......................................................... 30  Figure 16: Number of Kindergartens per 1000 Children ......................................................................... 30  Figure 17: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Kazakhstan 2017 ............ 51  Figure 18: Share of the population in the middle class, rates for Kazakhstan 2017 .............................. 52  Figure 19: Average Per Capita Daily Consumption in 2011 PPP for Kazakhstan 2017 ......................... 53  Figure 20: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Uzbekistan 2018 ............. 54  Figure 21: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Uzbekistan 2018 ............. 55  Figure 22: Share of population in the middle class, rates for Uzbekistan 2018 ..................................... 56  Figure 23: Average Per Capita Daily Consumption in 2011 PPP for Uzbekistan 2018 ......................... 57  Figure 24: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Kyrgyzstan 2016 ............. 58  Figure 25: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Kyrgyzstan 2016 ............. 59  Figure 26: Share of the population in the middle class, rates for Kyrgyzstan 2016 ............................... 60  Figure 27: Average Per Capita Daily Consumption in 2011 PPP for Kyrgyzstan 2016 .......................... 61  Figure 28: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Tajikistan 2015 ............... 62  Figure 29: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Tajikistan 2015 ............... 63  Figure 30: Share of the population in the middle class, rates for Tajikistan 2015 ................................. 64  Figure 31: Average Per Capita Daily Consumption in 2011 PPP for Tajikistan 2015 ............................ 65  Figure 32: Upper and Lower-bound Confidence Intervals, Including Out-of-Sample Predictions .... 66  Figure 33: Comparison between ELL and FH ........................................................................................ 67  Figure 34: Root Mean Square Error Before and After FH Approach (Poverty Headcount Ratio $5.5/day - Left; Average per-capita consumption 2011 PPP – Right) ................................................... 68  Figure 35: Comparison of Individual Models vs. Pooled Model (Only in sample - Left; Including out of sample - Right) ...................................................................................................................................... 68  Figure 36: Left: Strict VIF Threshold (5) vs. Adopted Threshold (10); Right: No VIF Threshold vs. Adopted Threshold (1)) ............................................................................................................................. 69  Figure 37: Direct vs. Fay Herriot Estimates for in-sample Poverty $5.5day (left) and average consumption (right) .................................................................................................................................. 70  Figure 38: Improvement of Coefficient of Variation in average consumption ..................................... 71  5    Figure 39 comparison of variance estimates for average consumption ................................................. 72  Figure 40: Comparison of variance estimates for poverty at $5.5/day .................................................. 72  Figure 41: Poverty Headcount Ratio (5.5/day) for Taylor Linearized Method vs. HT Method (Left), Taylor Liberalized Method vs. Smoothed (Right) .................................................................................. 73  Figure 42: Lasso AIC vs. Stepwise ........................................................................................................... 76  Figure 43: Lasso EBIC vs. Stepwise  ........................................................................................................ 77  Figure 44: Lasso BIC vs. Stepwise ........................................................................................................... 77  Figure 45: Fold A Full Sample (Left); Withheld Sample (Right) ........................................................... 78  Figure 46: Fold B Full Sample (Left); Withheld Sample (Right) ........................................................... 79  Figure 47: Fold C Full Sample (Left); Withheld Sample (Right) ........................................................... 79  Figure 48: Fold D Full Sample (Left); Withheld Sample (Right)  ........................................................... 80  Figure 49: Fold E Full Sample (Left); Withheld Sample (Right) ........................................................... 80  6    I – Introduction Measurement of poverty rates, average per capita consumption, the size of the middle class, and other welfare indicators is traditionally conducted using survey data. 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. This study starts with nationally (and regionally) representative data from each of the four Central Asian countries (Kazakhstan, Kyrgyzstan, Tajikistan and Uzbekistan) for which they are available. The analysis then proceeds to sharpen the reliability of the survey estimates to allow reporting at a level below what is traditionally reported (moving from “oblast” level estimates, to “rayon” level estimates). Because no such survey data are available for Turkmenistan, the country is excluded from the analysis. There are two survey types used in the approach described in this report. In Kazakhstan and the Kyrgyz Republic, welfare estimates are derived from the Household Budget Surveys (HBS) conducted by the national statistical agencies. In Tajikistan and Uzbekistan, HBS data were either inappropriate for this analysis, or unavailable. Instead, the welfare estimates for these countries are derived from the baseline surveys for two World Bank studies: Listening to Tajikistan (L2T) and Listening to the Citizens of Uzbekistan (L2CU). Each of these four surveys are designed to be representative at the national level in the respective country, for rural and urban areas, and at the region (oblast) level. For this reason, the national official rates of poverty monitored in each country are not usually published for administrative units below the regional level (with some exceptions). However, many government administrative activities are undertaken at the district level, and regions are a relatively high level of spatial aggregation for targeting policies that are sensitive to the needs of poor and vulnerable people. Likewise, policies that are intended to grow the middle class and average income could benefit from greater disaggregation than is directly available. Small area estimation (SAE) techniques sharpen the precision of welfare measures to enable reporting at highly disaggregated geographic units. Using statistical models for imputation, SAE approaches provide estimates of indicators for small areas that would be impossible to reliably construct with traditional survey data alone. 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. In many countries and regions, poverty maps have been used to highlight geographic variations, simultaneously display different dimensions of well-being, understand income determinants, and to both design and select interventions. A variety of such methods have been devised to overcome the increasing imprecision of welfare estimates as they are disaggregated. The standard approach to SAE, used by the World Bank and applied in most cases when enough data are available, is described in Elbers, Lanjouw, and Lanjouw 7    (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. In Central Asia, census data either do not exist (for instance, in Uzbekistan), are relatively dated (for instance, in the Kyrgyz Republic and Tajikistan), or are not made available for these purposes (for instance, in Kazakhstan). In such cases, the most common alternative approach is the Fay-Herriot (FH) method, which is adopted to generate the results described in this report. The FH method allows estimation of indicators and rates using a combination of survey data and district-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 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 district, 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 district level (in this case, either the poverty rate, average consumption per capita, or the share of the middle class at the district level). Once the model is estimated, the direct survey estimates also enter into the final area-level results: the final estimated area-level poverty rate 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, the results rely entirely on estimates derived from the statistical model. The report is structured as follows. The remainder of this section reviews the definitions of the key indicators to be estimated. Section II reviews the data used in the study (relying on both survey microdata, and complementary satellite/administrative data). The approach used for the maps described in this report is discussed in more detail in section III. Section IV includes the district-level maps for Central Asia. Section V briefly discusses the results and gives two examples of uses for the maps (comparing migration rates to poverty rates and locating program locations on maps). The annexes include detailed tables of the results at the level of districts (rayons) throughout Central Asia, validation and robustness exercises, and the regression models used. I.I – Definitions of Poverty in Central Asia The World Bank regularly produces internationally comparable estimates of poverty as part of its mandate. In 2017, the World Bank updated all such estimates to a new set of poverty lines, and at the same time, began applying up-to-date purchasing power parity (PPP) conversion factors for each country (based on estimates of price differences in 2011). The objective of calculating the World Bank’s internationally comparable poverty estimates is to estimate the prevalence of poverty in terms of a single global standard. Measures created on this basis are in turn used to monitor progress towards development goals set by the World Bank, the member states of the United Nations, and other partners. The World Bank poverty estimates contrast with national official poverty estimates in Central Asia in several ways. The most common official poverty approaches in the region consider local patterns of 8    income or consumption and are often more appropriate for country-specific analysis. However, despite their many advantages, national official poverty measures are not comparable with the approaches used in other counties, in part because they are so tailored the specific country context. Thus, comparisons across countries (for instance, for the purposes of monitoring or benchmarking) require a harmonized approach, such as that conducted by the World Bank. An additional reason that the World Bank uses a stand-alone approach to measure internationally comparable poverty rates is to account for differences in the cost of living between countries. To address this issue, the World Bank participates in the International Comparisons Program (ICP), a global effort to measure differences in the amount of goods and services a unit of one country’s currency can purchase in another country. The ICP exercise was most recently completed in 2011, which led to important updates to previous PPP measures. The original dollar-a-day line was created in 1991 using 1985 PPPs and was estimated by taking an average of the national poverty lines in the world’s poorest countries. When a new set of PPPs was published in 1993, the line changed to $1.08 per day. PPPs were revised again in 2005, and the new line was estimated at $1.25.2 For the most recent revision in 2011, the new line was estimated again using the national poverty lines of the poorest countries, as was the case in previous rounds. The result was a line set at $1.90 a-day. However, poverty by this measure is quite rare in Central Asia and is therefore excluded from the analysis that follows. But in addition to the international “low income” poverty line, the World Bank also uses income class poverty lines which facilitate comparisons between countries at similar stages of development. The income class poverty lines are defined for the lower middle-income and upper middle-income countries and are based on the national poverty lines of the countries in each group. As such, they provide a more appropriate threshold to measure poverty for countries in each income class. The lines are defined at $3.2 (for lower middle-income countries) and $5.5 (for upper middle-income countries). The welfare measures of income or consumption used are the same as those used for the international poverty line. Cross-country comparisons of poverty measured at these lines are presented in Figure 1.                                                              2 For more details on the updated lines from Francisco Ferreira, Dean Joliffe and Espen Prydz, follow this link 9    Figure 1: Cross-Country Comparison of Income Class Poverty Rates 2002/3 2014/15 100 90 80 Poverty Headcount 70 60 50 40 30 20 10 0 Turkey Turkey Turkey Armenia Georgia Kazakhstan Armenia Georgia Kazakhstan Armenia Georgia Kazakhstan Tajikistan Ukraine Tajikistan Ukraine Tajikistan Ukraine Kyrgyz Republic Russian Federation Kyrgyz Republic Russian Federation Kyrgyz Republic Russian Federation international poverty line lower middle income class poverty upper middle income class poverty line line Source: Official Survey Data by Country, Author’s calculations Kazakhstan has achieved large and enduring reductions in poverty since independence. The rate of extreme poverty measured at the international poverty line of $1.9-a-day is now largely indistinguishable from zero. Despite setbacks in 2009 and 2015, poverty measured using all common lines indicate substantially improving welfare over the long-term. The most recent data available suggest that the poverty rate in Kazakhstan fell between 2016 and 2017. Measured against the highest poverty line the World Bank uses – which is usually reserved for upper middle-income countries like Kazakhstan – poverty has fallen from about 65 percent in 2011 down to about 8.5 percent in 2017. But many challenges remain, and Kazakhstan is vulnerable to economic shocks that periodically reverse the country’s progress on poverty reduction. The two most notable cases of this were in 2005 and from about 2013 to 2016. During those spells of low growth or contraction, many people fell into poverty (though not often extreme poverty), and only a return to growth in the following years pulled the poverty rate down again. Indeed, measured at the highest poverty line, Kazakhstan has still yet to return to the low rate it achieved in 2013. Even at the regional level, welfare and resilience to shocks have strong spatial dimensions in Kazakhstan. During the most recent economic downturn starting in late 2014 and continuing into 2015, regions with lower average incomes and histories of higher poverty rates experienced larger increases in poverty in 2015. 10    Figure 2: Poverty rates in Kazakhstan using international poverty lines 70.0% 60.0% 50.0% 40.0% 30.0% 20.0% 10.0% 0.0% 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 2012 2013 2014 2015 2016 2017 $1.9 PPP $3.2 PPP $5.5 PPP Source: the Kazakhstan HBS, Author’s calculations Among the Central Asian countries, Tajikistan traditionally has the highest rates of poverty, though in recent years the gap with Kyrgyzstan has been falling. As is the case with Kazakhstan, the country monitors poverty using a national official definition that is distinct from the World Bank’s poverty line. According to the national series, the poverty rate fell from over 72 percent in 2003 to 47 percent in 2009. The rate of extreme poverty measured at the international poverty line of $1.9-a-day is below 2 percent. According to a different method of measuring national poverty which began in 2012, poverty continued to decline from 37.4 percent in 2012 to 29 percent in 2017. The poverty rate in Tajikistan fluctuates strongly between seasons and about half of the population falls into poverty (by the national definition) at least once during the year (figure 3). 11    Figure 3: National Poverty Rates in Tajikistan (left-note not comparable to international poverty estimates); Number of quarters in poverty (right) 45% Number of Quarters a Household is  40% Poor During the Year 35% 100% 90% 30% 80% 25% 70% 20% 60% 15% 50% 10% 40% 30% 5% 20% 10% 0% 2013 2014 2015 2016 2017 Poverty Ext. Poverty No Poor Qtr One Poor Qtr Two Poor Qtr Qtr. Poverty Qtr. Ext. Poverty Three Poor Qtr Four Poor Qtr Source: the Tajikistan HBS, Author’s calculations Kyrgyzstan struggles with poverty rates nearly as high as those in Tajikistan, and in some cases higher. Although poverty fell rapidly for the first half of the past decade, progress stagnated thereafter. Measured at $3.2-a-day, the poverty rate fell from about 79 percent of the population in 2000 to 20 percent in 2008. However, the declining trend nearly halted at that point, and over the following years has remained nearly flat. Changes in the poverty rate are strongly associated with economic growth in the Kyrgyzstan, and recent slow economic growth has translated into little poverty reduction. The country is also relatively undiversified, and in the past, boom-bust cycles in economic activity have been closely linked to periods of progress and stagnation in poverty reduction. 12    Figure 4: Poverty rates in the Kyrgyz Republic Source: the Kyrgyzstan HBS, Official World Bank calculations Until recently, Uzbekistan did not regularly provide the international community with the official data needed to estimate internationally comparable poverty rates. Consequently, the latest official and internationally comparable estimates date from the early 2000s. However, in 2018 a new study was launched by the World Bank in consultation with the National Statistical Agency of Uzbekistan and other partners called Listening to the Citizens of Uzbekistan. This study included a comprehensive baseline survey that can be used to estimate comparable poverty rates. These estimates suggest that in 2018 the poverty rate measured at the $3.2/day line stood at 9.6 percent of the population, and 36.6 percent at the $5.5-a day line. 13    Figure 5: Poverty Rates by Region of Uzbekistan (2018) 40% 35% 30% 25% 20% 15% 10% 5% 0% $3.2/Day (Inc) $3.2/Day (Cons) Source: The Listening to the Citizens of Uzbekistan baseline survey, Author’s calculations Turkmenistan alone has no data available at all to estimate internationally comparable estimates. For this reason, it is excluded from the analyses that follow. In each of the four countries for which data are available, the survey data that are used to generate poverty estimates are usually considered to be representative at the level of regions (rayons) only. When disaggregated further, sampling errors are sufficiently large that confidence intervals substantially overlap. The resulting maps can potentially be misleading. Point estimates indeed suggest a specific value for the poverty rate in a given district, but if the precision of these estimates is ignored, it is easy to miss that the true value of the poverty rate may be much higher or lower. Such interpretations can lead to inappropriate policy decisions. Figure (6) presents harmonized poverty estimates for the four countries in which survey data are available at the level that the underlying surveys are usually considered to be representative. The task for the remainder of this report is to estimate and report welfare indicators, including poverty rates, that are representative at the district level, one step lower than those presented here. 14    Figure 6: UMIC Poverty Rate by Regions (Oblasts) in Central Asia Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. I.II – The Middle Class in Kazakhstan, and Average Per Capita Consumption In Kazakhstan and in each country in Central Asia, growing the middle class is a central aspiration of the government, and one of the key commitments laid out in National Development Strategies. Although poverty has fallen, many people remain vulnerable to economic volatility. A large share of population has a consumption level just above the poverty line for upper middle-income countries and at significant risk of falling below it. As the recent economic downturn in Central Asia showed, there is substantial churning in and out of poverty among households during times of economic hardship, which points to a low degree of resilience. In the context of increased economic volatility, region-wide or idiosyncratic shocks easily translate to reduction in welfare. From an economic perspective, exiting that state of vulnerability could be viewed as a transition process to becoming a “middle class” society and highlights the process of reaching a threshold of economic stability associated with a low probability of falling back into poverty. There currently is no official measure of the middle class in Kazakhstan or in the other countries in the region, but in Pittau 15    & Zelli (2018), the World Bank together with other development partners recently recommended a strategy based on data from the official HBS. Fixed thresholds are a common approach to measuring and monitoring the evolution of the middle class. But there is a wide variety of thresholds that have been used in varying contexts. One of the most replicated comes from Ferreira et al. (2013) who estimated per capita middle-class income lines of US$10 to US$50 a day in 2005 PPP terms in Latin America and the Caribbean. However, it is not clear that this range makes sense for Kazakhstan, which is quite removed from the context in which these lines were initially set. Thus, the method recommended by the World Bank, and adopted for this study, uses an empirically estimated line calibrated to the context of Kazakhstan using a module of self-reported class membership in the 2013 Kazakh HBS. The approach yields a lower income threshold of 474,000 tenge and an upper threshold of 1,772,000 tenge, which correspond to the 56th and the 99th percentiles of the weighted income distribution. These values in turn correspond to US$14.00 and US$52.20 per day in 2011 international dollars. Because at this time there is no official middle-class line for the other countries in the region, and the line that has been developed in Kazakhstan is the most articulated definition in the region, this line is used throughout the region for the maps that follow. Figure 7: The middle class share of the population, by regions (oblasts) of Central Asia Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. Another important measure of welfare is average consumption, which is a strong proxy for income in countries with large informal and agriculture sectors. Average consumption focuses attention not only 16    on the bottom of the distribution such as poverty indicators in this context. Average consumption can also be measured more precisely and is amenable to small-area estimates in ways that poverty is not in most applications. One drawback of the Fay-Herriot approach in the context of values close to zero (such as percentage values) is that they are more difficult to validate in terms of coefficients of variation (National Research Council, 2013). However, average consumption suffers from no such limitations. To be comparable across countries, average consumption is adjusted using the PPP conversion factor for each country and expressed in 2011 terms. Figure (8) presents the direct estimates at the regional level. Figure 8: Average Daily Per Capita Consumption in 2011 PPP Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. II – Data Data from two types of sources are generally required to conduct an SAE mapping exercise. The first source is a welfare survey, which is preferably the data with which poverty, consumption, and/or the size of middle class is monitored. The second source must permit disaggregation to the level for which the desired indicator will be estimated and, preferably, include the entire population, rather than a 17    sample. Any sampling for the second source leads to additional error and should be avoided if possible. For the area-based estimates described in this report, publicly available satellite, station, and geographic data aggregated to the district-level are used, none of which suffers from additional constraints due to sampling. To ensure that the definitions of key indicators defined in the microdata are as similar across countries as possible, the harmonization standard applied by the World Bank in its ECAPOV initiative in generating official World Bank estimates of poverty is used here. II.I – Districts by Country Presenting the results and matching the survey data to satellite imagery require identifying the cartography of each country at the district level. This was accomplished using standardized map shapefiles and matching each to the disaggregated location data available for each survey. Table 1: Survey Data Used Svy Map Country Year Hhlds Dist. Dist. Impute KAZ 2017 47445 178 202 24 KGZ 2016 19575 40 42 2 TJK 2015 2999 149 159 10 UZB 2018 4013 126 190 64 TKM None . . . . The analyses discussed in this report primarily use the data from the diary/list of food consumed, non-food expenditure, and household composition. II.II – Kazakhstan HBS data for 2017 Poverty estimates and other welfare indicators for Kazakhstan rely on the 2017 round of the HBS, conducted by the Statistics Agency of the Republic of Kazakhstan. For household consumption and expenditure, the survey is usually considered nationally representative, representative at the oblast (region) level, and separately representative for rural and urban areas. The survey uses a stratified sample design with strata corresponding to 14 regions + 2 large cities crossed by their urban and rural areas (except for Almaty and Astana cities, which are entirely urban). A complete consumption module is gathered in the HBS, covering both food and non-food items. Information on household composition, income, employment, and related topics is also collected. All data for the HBS in Kazakhstan are collected in person, except for consumption, which is collected using a diary-type instrument. The survey is continuous, and representative quarterly within the year for most welfare indicators. 18    II.III – Uzbekistan L2CU Baseline Survey 2018 Official HBS data for Uzbekistan had not been made available by the National Statistical Agency to the World Bank until recently. However, the L2CU activity, which is conducted by a private firm on behalf of the World Bank and in cooperation with Ministries and the Statistical Agency, included a comprehensive baseline survey appropriate to the needs of poverty mapping. The survey design closely followed the Living Standards Measurement Study (LSMS) surveys and was conducted using a standard two-stage sampling design, in which 200 clusters (administrative territories called mahallas, in the case of Uzbekistan) were randomly selected with probability proportionate to size. The national sample was stratified by region and by urban areas. Though a national household census would be a preferable sample frame for the first selection step; no such census has been conducted in Uzbekistan since independence from the Soviet Union (a census is currently scheduled for 2022). Instead, the universe of mahallas, the smallest official jurisdiction in Uzbekistan, was used. These data include official population statistics provided by the National Mahalla Committee from official registration data (current as of the first quarter of 2018). The data were re-weighted based on observed population totals within the mahallas at the time of the survey fieldwork. The second stage procedure was conducted using simple random selection with equal probability within selected mahallas. 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 cross- referencing and other techniques to ensure accuracy. The survey was conducted over the course of a 1.5-month period in May/June 2018. Although the L2CU study collected data continuously, the baseline is the only source that provides a comprehensive consumption and expenditure module. II.IV – Kyrgyz Republic HBS 2016 Poverty rates and other welfare estimates for Kyrgyzstan rely on the 2016 round of the Kyrgyzstan HBS, conducted by the National Statistical Agency. For household consumption, the survey is usually considered nationally representative, representative at the oblast (region) level, and separately representative for rural and urban areas. The survey uses a stratified sample design with strata corresponding to seven regions + 1 large city (Bishkek) crossed by their urban and rural areas. A complete consumption module (using a diary approach) is gathered in the HBS, covering both food and non-food items. Information on household composition, income, employment, and related topics is also collected. All data are collected in-person, and in most recent years using a CAPI system. 19    II.V – Tajikistan L2T Baseline Survey for 2015 Although the Tajikistan HBS survey is made available to the World Bank and other development partners, the sampling design used is not well-suited to the purposes of poverty mapping because the geographic coverage of the survey is limited in comparison to other options. Geographic coverage is crucial in the context of poverty mapping, and as an alternative, a survey with a more widely distributed sample was used. The baseline for the L2T Survey was selected because it: i) used the 2010 national census sample frame for the L2T baseline survey conducted by the national statistical agency, and ii) used a widely distributed traditional stratified two-stage clustered sample design. In the first stage, 150 clusters were selected, with a probability of selection proportional to size. In the second stage, 3,000 households were selected to participate in the survey. The sample was designed to be nationally representative for consumption and expenditure. The survey was conducted over a two-month period beginning in March 2015. CAPI systems were used and all data were collected in person. As with the L2CU study, continuous data collection was conducted following the baseline survey, however, no comprehensive consumption or expenditure model was included in follow-up rounds. The interviews were implemented under the supervision of the Ministry of Health and Social Protection of Tajikistan and the World Bank. II.VI – Satellite and Administrative Data The main source of satellite and administrative data for this study is the AidData project at the college of William and Mary.3 The spatially aggregated data available for Kazakhstan from this source include: 1) World Bank project locations 2) Yearly VIIRS day night band nighttime lights data (without stray light correction) 3) Version 4 DMSP-OLS Nighttime Lights composites. The lights from cities, towns, and other sites with persistent lighting, including gas flares. Ephemeral events, such as fires have been discarded. Calibrated across sensors and years using Elvidge 2014 coefficients 4) Average precipitation per year, created using UDel Precipitation data set (v4.01) 5) Average air temperature per year, created using UDel Air Temperature dataset (v4.01) 6) Global slope (in degrees) derived from Shuttle Radar Topography Mission (SRTM) data set (v4.1) at 500m resolution 7) Global elevation (in meters) from Shuttle Radar Topography Mission (SRTM) data set (v4.1) at 500m resolution 8) Binary indicating locations with deposits of known on-shore oil and gas deposits 9) Standard MODIS land cover type data product (MCD12Q1) in the IGBP Land Cover Type Classification 10) Yearly value for Normalized Difference Vegetation Index (NDVI). Created using the NASA Long Term Data Record (v4) AVHRR data                                                              3 Data available at http://geo.aiddata.org/query/#!/ 20    11) Population density (UN Adjusted values) from Gridded Population of the World v4.GPWv4 depicts the density of human population across the globe; source data provided in 30 arc- second (~1 km) grid cells 12) Map of total economic activity, including both formal and informal economic activity for ~2006; created from nighttime lights and LandScan population grid 13) Distance to coast (on land only), measured in meters; derived using World Vector Shorelines 14) Distance to water, measured in meters; derived using World Vector Shorelines combined with rivers and lakes from World Data Bank 2 (via Natural Earth) 15) Distance to roads, measured in meters, based on the Global Roads Open Access Dataset (gRoads) version 1.0 16) Distance to country borders, measured in meters; derived using GADM 2.8 ADM0 (country) boundaries 17) Particulate matter (PM2.5) estimate, based on prediction model using combination of satellite- based estimate and TM5-FASST simulation 18) Estimated travel time (in minutes) to the nearest city of 50,000 or more people in year 2000. III – Fay-Herriot Small Area Estimation Model The basic area-level model setup is as follows. Let be the true poverty or middle-class incidence in each geographic area i, and let the sampling model be defined by: , where is the observed survey direct estimate of poverty headcount ratio (or middle-class headcount ratio) , 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 . The data on are obtained from publicly available sources at the area-level, largely from administrative sources focusing on geography, and from continuous monitoring data from satellites or weather stations, and hence are free of sampling error. Combining the above sampling and linking models, it follows that the observed poverty or middle-class membership rates from the survey can be modeled as follows: . Given this setup, the best linear unbiased estimator of , one that minimizes the mean squared error is: , 21    where , and is referred to as a “shrinkage factor”. Given that is unknown, the Best Linear Unbiased Predictor (BLUP) is replaced with its empirical counterpart EBLUP: , which can be rewritten as: 1 , where is the Feasible GLS estimator for and . 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 . The prediction error associated with ̂ takes account of the sampling variance associated with , as well as the uncertainty associated with the estimate of and (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. Perhaps the simplest approach is to use the Horvitz-Thompson (HT) estimator of , which can be expressed as: ∈ The term denotes the set of sampled units in domain d. However, the HT estimator is often not preferred in SAE applications (National Research Council, 2013; Eurostat, 2013). It is commonly unstable, especially for small domains with few survey observations (a common challenge in this context). To address this, there are various alternative direct estimators that may out-perform the HT estimator. Common practice in the World Bank has been to obtain sampling variances associated with the district-level poverty rates 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 simple weighted direct HT estimates. A final 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 of the three described here, and the preferred approach for this application. Annex G reports comparisons and analysis of sensitivity to the choice of domain variance measure. Final results are quite similar when comparing the second (“smoothed”) and third (“linearized”) options described, though as expected, the HT approach yields results that are more distant from those of the remaining two options. For more detail on the trade- off between approaches for domain variance estimation, see Heeringa et. al., (2017); Molina and Rao (2010); Wolter (2007); and Kolenikov (2010). 22    The results of the SAE estimates are presented graphically in the following section. Separate modeling exercises are conducted for each welfare indicator. 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 R2.4 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. Sensitivity analysis comparing the results to alternative model development approaches (including several variations of the lasso approach) are described in annex I. IV – Results Summarizing poverty rates by region (at which the underlying surveys are themselves considered representative) yields a range of between 0 and 27.7 percent at the poverty line for lower middle- income countries ($3.2 PPP per person, per day), and between 0 and 76 percent for the line appropriate for upper middle-income countries ($5.5 PPP per person, per day). But this level of aggregation hides significant variation within regions. When disaggregated to the district level, the resulting poverty estimates for the LMIC line range from 0 percent to 70 percent, and many individual districts register nearly 100 percent poverty rates at the UMIC line, despite a lower average poverty rate in the surrounding region.5 Poverty rates at the LMIC line are relatively low throughout Kazakhstan and in much of Uzbekistan. However, pockets of high poverty rates remain in the Kyrgyz Republic and in Tajikistan. The parts of the Ferghana Valley within Uzbekistan’s borders have relatively low rates of poverty but nearby districts in the Kyrgyz Republic and Tajikistan with somewhat higher rates of poverty. Even stronger regional disparities are apparent when using the UMIC line. Poverty by this definition was still less than 10 percent in Kazakhstan in 2017, but this stands in contrast to much higher rates in the Kyrgyz Republic, Tajikistan, and to a lesser extent, Uzbekistan. In Tajikistan and the Kyrgyz Republic there are districts in which nearly all people are poor by the UMIC line, whereas the highest rate in Kazakhstan is 44 percent. Throughout Central Asia, most of the poor live in the most populated districts, despite the higher poverty rates prevalent in some of the sparsely populated and mountainous regions. The density of the population below the poverty threshold (i.e. the absolute number of poor individuals, obtained as the product of the predicted poverty rate and population of the district) is concentrated in the more densely populated areas in each country. Rural areas are much poorer than urban areas on average, and remote areas have higher average poverty rates. Districts that border on Afghanistan also tend to be poorer. Figures 9-12 (and 17-31) in the following depict each of the welfare indicators derived using the Fay-Herriot approach in this application.                                                              4  Adjusted R is chosen instead of (unadjusted) R because the latter is non-decreasing in the number of explanatory variables in the model. 5 The district-level estimates are reported in Annexes A-D.  23    Figure 9: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Central Asia Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. 24    Figure 10: Poverty at $5.5-a-day PPP per person, rates for Central Asia Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. 25    Figure 11: Average Per Capita Consumption 2011 PPP Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. 26    Figure 12: Share of Population in the Middle Class in Central Asia Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). All welfare measures are reported in real PPP terms. Welfare data are spatially deflated within country. 27    V – Concluding Remarks and Example Applications Millions of people have left poverty in Central Asia over the past two decades. But some areas are improving more quickly than others. Insufficient sample sizes in most standard surveys are a barrier to better understanding the dynamics of economic convergence in Central Asia. This study partially addresses the challenge. Key welfare indicators are improved upon by unifying survey data from across the region and applying the techniques of small-area estimation. These maps have many potential practical uses. Given the high rates of migration and remittance- dependence in the region, one important example is to compare rates of migration to poverty rates at the district level. Figure (13) provides the share of households with at least one migrant member by district. Figure 13: Share of Households with at Least One Member Abroad for Work Notes: Sources include the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). Comparing these estimates highlights how labor migrants from Central Asia disproportionally originate in some of the poorest districts in the region. The correspondence suggests that migrants significantly contribute to the region’s poorest areas through both remittances and investment. About 19 percent of households in bottom quintile districts have at least one migrant abroad, compared to less than 2 percent in top quintile districts. 28    Figure 14: District Share of HHs with Migrants by District Per Capita Consumption Quintile 25% 20% 15% 10% 5% 0% Bottom Quint 2 3 4 Top Quint Share of HHs w/Migrants Source: Author’s calculations on the basis of the baseline survey for Listening to the Citizens of Uzbekistan (2018), the baseline survey for Listening to Tajikistan (2015), the household budget survey of Kazakhstan (2017), and the household budget survey for Kyrgyzstan (2016). These results complement other analyses that focus on migration; providing useful regional context to this analysis. Where household-level determinants of migration have been analyzed in more detail, there is a strong relationship between migration and local economic challenges, including low labor force participation, household welfare shocks, and living in an area with greater dependence on social protection benefits. Absent financial support from migrants, poverty rates and unemployment rates in struggling regions would be significantly higher, while average incomes would be much lower. Another example is country specific: in Tajikistan, the government and World Bank are in advanced discussions of expanding the number of kindergartens and early childhood learning centers. But what areas need more assistance? The following figures (15-16) overlay the poverty maps for the country with details on the number of educational facilities per 1,000 children, highlighting areas with significant co-incidence of high poverty rates and low coverage. 29    Figure 15: Number of Early Learning Centers per 1000 Children Source: Author’s calculations based on administrative data from the Ministry of Education in Tajikistan and the baseline survey for Listening to Tajikistan (2015) Figure 16: Number of Kindergartens per 1000 Children Source: Author’s calculations based on administrative data from the Ministry of Education in Tajikistan and the baseline survey for Listening to Tajikistan (2015) 30    These results thus provide detailed measures of welfare that in turn can be disaggregated for each district in Central Asia. The results highlight the heterogeneity of progress in the region. Poverty estimates range from 0 to nearly 100 percent by both the LMIC poverty line and the UMIC line. The results are robust to many separate validation approaches. The maps created with these data can be used to identify areas where progress is being achieved most rapidly, as well as those areas that are lagging. With better and more detailed information at the ready, policies and interventions can be better targeted to support people in need where they live. 31    VI – References Ahrens, A., Hansen, C.B., Schaffer, M.E. 2018. lasso2: Program for lasso, square-root lasso, elastic net, ridge, adaptive lasso and post-estimation OLS. Bedi, T., A. Coudouel and K. Simler, (2007) More than a pretty picture: using poverty maps to design better policies and interventions. Washington DC: The World Bank Group. Elbers C., J. O. Lanjouw and P. Lanjouw (2003) “Micro-Level Estimation of Poverty and Inequality”. Econometrica 71(1): 355–364. Fay, R. and R. Herriot. (1979) “Estimates of income for small places: an application of James-Stein procedures to census data.” Journal of the American Statistical Association 74 (1979): 269–277. Goodman, S., Ben Yishay, A., Runfola, D., 2016. Overview of the geo Framework. AidData. Available online at geo.aiddata.org. DOI: 10.13140/RG.2.2.28363.59686 Heeringa, Steven G., Brady T. West, and Patricia A. Berglund. Applied survey data analysis. Chapman and Hall/CRC, 2017. Molina I. and Rao J. (2010) “Small area estimation of poverty indicators.” Canadian Journal of Statistics, 38(3), 369-385. National Research Council. National Patterns of R&D Resources: Future Directions for Content and Methods: Summary of a Workshop. National Academies Press, 2013. Office for Official Publications of the European Communities, Luxembourg. Eurostat (2013). Handbook on precision requirements and variance estimation for ESS household surveys Rao, J. N. K. (2003) Small Area Estimation. 1st ed. Wiley-Interscience. Statistical Office of the Republic of Kazakhstan (2015) “Income and Living conditions in the Republic of Kazakhstan – 2013”. Final Report. Belgrade, Republic of Kazakhstan: Statistical Office Kolenikov, Stanislav. "Resampling variance estimation for complex survey data." The Stata Journal 10.2 (2010): 165-199. Wolter, Kirk. Introduction to variance estimation. Springer Science & Business Media, 2007. World Bank (2012) “Pilot Study of Small Area Poverty Estimation Methods for the New Member States of the European Union,” Report prepared for the Scientific Steering Committee of the World Bank and European Commission Project on Small Area Poverty Estimation, Washington, DC: The World Bank Group. 32    VII – Annexes Annex A – Kazakhstan District-level Results 2017 Table 2: Kazakhstan District-level Results 2017 Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Almaty Aksuskiy 0.000 0.000 0.056 0.021 0.571 0.039 14.896 0.643 Almaty Alakolskiy 0.000 0.000 0.132 0.035 0.657 0.043 13.741 0.838 Almaty Almaty (Alma-Ata) 0.001 0.001 0.020 0.004 0.574 0.013 14.828 0.221 Almaty Balkhashskiy * 0.000 0.083 0.143 0.144 0.846 0.077 9.897 1.832 Almaty Enbekshikazakhskiy 0.000 0.000 0.092 0.013 0.777 0.018 11.332 0.303 Almaty Iliyskiy 0.000 0.000 0.020 0.010 0.728 0.029 12.578 0.722 Almaty Karasayskiy 0.000 0.000 0.048 0.009 0.610 0.019 14.202 0.338 Almaty Karatal`Skiy * 0.053 0.081 0.201 0.141 0.871 0.075 9.595 1.794 Almaty Kerbulakskiy * 0.044 0.082 0.125 0.142 0.821 0.076 10.706 1.801 Almaty Koksuskiy 0.000 0.000 0.024 0.015 0.639 0.043 14.114 0.915 Almaty Panfilovskiy 0.000 0.000 0.039 0.013 0.881 0.020 9.824 0.276 Almaty Raiymbekskiy 0.000 0.000 0.000 0.000 0.594 0.038 14.369 0.350 Almaty Sarkandskiy 0.000 0.082 0.043 0.142 0.784 0.076 11.538 1.811 Almaty Taldyqorghan 0.000 0.000 0.013 0.007 0.620 0.027 14.151 0.483 Almaty Talgarskiy 0.000 0.000 0.026 0.011 0.570 0.032 15.789 0.698 Almaty Uygurskiy 0.000 0.000 0.016 0.011 0.862 0.028 10.246 0.396 Almaty Zhambylskiy 0.000 0.000 0.020 0.011 0.680 0.033 12.491 0.511 Aqmola Akkol`Skiy 0.008 0.008 0.111 0.027 0.810 0.031 10.340 0.456 Aqmola Arshalynskiy 0.049 0.034 0.153 0.054 0.842 0.047 9.804 0.855 Aqmola Astrakhansiy 0.000 0.000 0.019 0.024 0.749 0.062 12.307 0.857 Aqmola Atbasarskiy 0.020 0.017 0.236 0.050 0.787 0.043 10.093 0.689 Aqmola Bulandynskiy 0.000 0.000 0.066 0.043 0.778 0.058 11.137 0.943 33    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Aqmola Egindykol`Skiy * 0.022 0.082 0.193 0.141 0.811 0.079 9.715 1.790 Aqmola Enbekshil`Derskiy * 0.019 0.081 0.123 0.141 0.823 0.079 10.239 1.792 Aqmola Ereymengauskiy 0.022 0.018 0.116 0.038 0.798 0.042 10.584 0.651 Aqmola Esil`Skiy 0.000 0.000 0.058 0.041 0.782 0.059 11.156 0.941 Aqmola Korgalzhynskiy * 0.008 0.082 0.175 0.141 0.783 0.080 9.842 1.791 Aqmola Sandyktauskiy 0.000 0.000 0.106 0.043 0.788 0.050 10.505 0.665 Aqmola Shortandinskiy 0.000 0.000 0.000 0.000 0.760 0.061 12.406 1.044 Aqmola Shuchinskiy 0.010 0.008 0.134 0.029 0.827 0.030 10.104 0.467 Aqmola Tselinogradskiy 0.000 0.000 0.015 0.004 0.681 0.014 12.983 0.166 Aqmola Zerendinskiy 0.010 0.007 0.111 0.023 0.739 0.030 11.893 0.514 Aqmola Zhaksynskiy 0.048 0.035 0.228 0.065 0.808 0.052 10.163 1.022 Aqmola Zharkainskiy 0.000 0.000 0.126 0.059 0.884 0.046 8.444 0.612 AqtöBe Alginskiy 0.000 0.000 0.084 0.053 0.891 0.052 9.661 0.761 AqtöBe Aqtobe 0.002 0.002 0.072 0.011 0.791 0.017 10.940 0.235 AqtöBe Aytekebiyskiy 0.007 0.012 0.107 0.043 0.868 0.044 10.050 0.887 AqtöBe Bayganinskiy 0.014 0.021 0.191 0.068 0.986 0.019 7.216 0.399 AqtöBe Irgizskiy 0.018 0.026 0.225 0.071 0.986 0.017 7.460 0.476 AqtöBe Kargalinskiy 0.000 0.000 0.000 0.000 0.812 0.065 12.026 1.072 AqtöBe Khobdinskiy 0.000 0.000 0.068 0.038 0.837 0.049 10.658 1.254 AqtöBe Khromtauskiy 0.000 0.000 0.090 0.039 0.852 0.042 10.173 0.645 AqtöBe Martukskiy 0.000 0.000 0.275 0.083 0.881 0.052 9.308 1.077 AqtöBe Mugalzharskiy 0.018 0.013 0.186 0.039 0.947 0.021 7.563 0.320 AqtöBe Shalkarskiy 0.000 0.000 0.133 0.043 0.878 0.038 9.500 0.536 AqtöBe Temirskiy 0.000 0.000 0.176 0.072 0.942 0.038 8.237 0.607 AqtöBe Uilskiy 0.000 0.000 0.000 0.000 0.792 0.062 12.295 1.017 Atyrau Atyrau 0.000 0.000 0.044 0.011 0.861 0.018 10.118 0.245 Atyrau Inderskiy 0.000 0.000 0.137 0.047 0.978 0.019 7.207 0.287 Atyrau Isatayskiy 0.000 0.000 0.078 0.036 0.910 0.035 8.853 0.513 34    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Atyrau Kurmangazinskiy 0.000 0.000 0.182 0.037 0.935 0.022 8.339 0.346 Atyrau Kzylkoginskiy 0.000 0.000 0.012 0.016 0.851 0.045 10.426 0.605 Atyrau Makatskiy * 0.022 0.081 0.146 0.141 0.921 0.076 9.540 1.800 Atyrau Makhambetskiy 0.000 0.000 0.049 0.033 0.842 0.047 10.796 0.643 Atyrau Zhylyoyskiy 0.000 0.000 0.000 0.000 0.759 0.040 12.393 0.538 East Ka Abayskiy * 0.010 0.081 0.143 0.141 0.888 0.076 10.045 1.794 East Kaz Ayagozskiy 0.098 0.040 0.344 0.062 0.873 0.041 8.913 0.933 East Kaz Beskaragayskiy * 0.006 0.081 0.143 0.141 0.873 0.075 10.074 1.790 East Kaz Borodulikhinskiy 0.000 0.000 0.109 0.035 0.818 0.040 10.704 0.637 East Kaz Glubokovskiy 0.000 0.000 0.044 0.026 0.704 0.046 12.924 0.904 East Kaz Katon-Karagayskiy 0.000 0.000 0.066 0.034 0.699 0.051 12.409 1.470 East Kaz Kokpektinskiy 0.000 0.000 0.048 0.031 0.908 0.035 9.349 0.461 East Kaz Kurchumskiy 0.013 0.016 0.201 0.057 0.811 0.047 10.762 0.920 East Kaz Leninogorsk 0.000 0.000 0.036 0.021 0.679 0.045 12.968 0.681 East Kaz Semipalatinskiy 0.003 0.003 0.050 0.011 0.721 0.021 12.106 0.313 East Kaz Shemonaikhinskiy 0.047 0.019 0.153 0.032 0.753 0.035 11.659 0.684 East Kaz Tarbagatayskiy 0.000 0.000 0.235 0.061 0.870 0.041 9.030 0.585 East Kaz Ulanskiy 0.004 0.003 0.045 0.010 0.577 0.023 15.149 0.500 East Kaz Urdzharskiy 0.008 0.010 0.192 0.042 0.895 0.029 8.548 0.404 East Kaz Zaysanskiy * 0.040 0.081 0.128 0.141 0.825 0.076 10.270 1.802 East Kaz Zharminskiy 0.000 0.000 0.026 0.021 0.895 0.038 9.180 0.550 East Kaz Zyryanovsk 0.000 0.000 0.019 0.014 0.644 0.042 13.886 0.793 Mangghystau Aqtau 0.000 0.000 0.012 0.007 0.901 0.019 10.202 0.200 Mangghystau Beyneuskiy 0.000 0.000 0.086 0.029 0.982 0.013 8.478 0.233 Mangghystau Karakiyanskiy 0.000 0.000 0.033 0.012 0.916 0.018 10.036 0.217 Mangghystau Manghystauskiy 0.000 0.000 0.015 0.012 0.900 0.029 10.194 0.337 Mangghystau Tupkaraganskiy 0.000 0.000 0.024 0.016 0.969 0.016 8.760 0.187 North Kaz Akzharskiy * 0.000 0.082 0.108 0.141 0.775 0.081 10.602 1.793 35    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. North Kaz Ayyrtauskiy 0.006 0.011 0.138 0.048 0.817 0.045 9.967 0.684 North Kaz Bulaevskiy 0.000 0.000 0.000 0.000 0.853 0.047 10.210 0.507 North Kaz Esil`Skiy 0.027 0.030 0.194 0.068 0.788 0.056 10.422 0.993 North Kaz Kyzylzharskiy 0.018 0.008 0.109 0.018 0.671 0.026 12.500 0.388 North Kaz Mamlyutskiy * 0.000 0.082 0.128 0.141 0.777 0.081 10.380 1.794 North Kaz Shal Akyna 0.009 0.011 0.236 0.048 0.779 0.042 10.346 0.745 North Kaz Sovetskiy 0.025 0.028 0.144 0.058 0.788 0.054 10.565 0.903 North Kaz Taiynshinskiy 0.000 0.000 0.000 0.000 0.744 0.058 12.593 0.645 North Kaz Timiryazevskiy 0.000 0.000 0.075 0.042 0.760 0.056 11.553 0.886 North Kaz Tselinniy 0.000 0.000 0.110 0.043 0.814 0.046 10.038 0.638 North Kaz Ualikhanovskiy * 0.000 0.082 0.098 0.141 0.764 0.082 10.502 1.794 North Kaz Zhambylskiy 0.000 0.000 0.019 0.023 0.788 0.055 11.125 0.804 Pavlodar Aksuskiy 0.000 0.000 0.030 0.016 0.722 0.038 12.534 0.530 Pavlodar Aktogayskiy * 0.000 0.082 0.131 0.141 0.808 0.081 10.092 1.793 Pavlodar Bayanaul`Skiy 0.000 0.000 0.080 0.043 0.804 0.052 10.150 0.657 Pavlodar Ekibastuz 0.000 0.000 0.061 0.019 0.716 0.032 11.743 0.414 Pavlodar Irtyshskiy 0.000 0.000 0.000 0.000 0.748 0.066 11.877 0.994 Pavlodar Kachirskiy 0.011 0.025 0.152 0.074 0.803 0.060 10.242 1.061 Pavlodar Lebyazhinskiy 0.000 0.000 0.176 0.078 0.809 0.062 10.180 1.073 Pavlodar Mayskiy 0.000 0.000 0.032 0.036 0.899 0.045 8.921 0.621 Pavlodar Pavlodarskiy 0.001 0.001 0.044 0.010 0.764 0.020 11.590 0.305 Pavlodar Sherbaktinskiy 0.000 0.000 0.041 0.035 0.800 0.057 10.568 0.837 Pavlodar Uspenskiy 0.014 0.030 0.116 0.069 0.723 0.069 11.141 1.546 Pavlodar Zhelezinskiy 0.007 0.015 0.081 0.045 0.733 0.059 11.939 0.844 Qaraghandy Abayskiy 0.007 0.006 0.095 0.022 0.671 0.033 13.244 0.607 Qaraghandy Aktogayskiy 0.000 0.000 0.075 0.030 0.775 0.041 11.383 0.618 Qaraghandy Bukhar-Zhyrauskiy 0.001 0.001 0.055 0.008 0.652 0.016 13.576 0.278 Qaraghandy Karkaralinskiy 0.000 0.000 0.081 0.046 0.839 0.049 9.695 0.706 36    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Qaraghandy Nurinskiy 0.010 0.020 0.104 0.058 0.777 0.062 11.091 1.346 Qaraghandy Osakarovskiy 0.000 0.000 0.137 0.046 0.732 0.052 11.835 0.872 Qaraghandy Shetskiy 0.011 0.012 0.179 0.044 0.878 0.033 8.692 0.446 Qaraghandy Ulytauskiy 0.018 0.008 0.152 0.021 0.833 0.022 10.363 0.592 Qaraghandy Zhanaarkinskiy 0.023 0.025 0.171 0.059 0.810 0.050 9.690 0.709 Qostanay Altynsarinskiy 0.013 0.022 0.161 0.062 0.842 0.051 9.348 0.728 Qostanay Amangel`Dinskiy * 0.000 0.081 0.127 0.141 0.912 0.076 9.746 1.798 Qostanay Arkalyk 0.000 0.000 0.000 0.000 0.897 0.037 10.732 0.377 Qostanay Auliekol`Skiy 0.000 0.000 0.145 0.042 0.853 0.038 9.792 0.588 Qostanay Denisovskiy 0.000 0.000 0.063 0.038 0.793 0.055 12.289 1.023 Qostanay Dzhangil`Dinskiy * 0.000 0.081 0.128 0.141 0.915 0.076 9.445 1.808 Qostanay Fyodorovskiy 0.000 0.000 0.198 0.070 0.828 0.055 10.047 0.941 Qostanay Kamystinskiy 0.053 0.036 0.237 0.065 0.915 0.039 8.195 0.579 Qostanay Karabalykskiy 0.037 0.031 0.184 0.060 0.929 0.034 8.596 0.698 Qostanay Karasuskiy 0.000 0.000 0.000 0.000 0.887 0.044 10.149 0.573 Qostanay Mendykarinskiy 0.000 0.000 0.102 0.039 0.807 0.045 11.066 0.718 Qostanay Naurzumskiy * 0.000 0.081 0.128 0.141 0.896 0.075 10.052 1.792 Qostanay Qostanay 0.000 0.000 0.030 0.008 0.757 0.019 11.501 0.259 Qostanay Sarykol`Skiy 0.000 0.000 0.124 0.050 0.962 0.024 8.185 0.399 Qostanay Taranovskiy 0.000 0.000 0.190 0.043 0.904 0.029 8.755 0.450 Qostanay Uzunkol`Skiy 0.000 0.000 0.000 0.000 0.863 0.048 10.605 0.557 Qostanay Zhitikarinskiy 0.000 0.000 0.000 0.000 0.783 0.052 12.712 0.763 Qyzylorda Aral`Skiy 0.000 0.000 0.080 0.025 0.884 0.026 9.547 0.480 Qyzylorda Karmakchinskiy 0.000 0.000 0.000 0.000 0.907 0.042 9.823 0.543 Qyzylorda Kazalinskiy 0.000 0.000 0.162 0.039 0.903 0.031 8.920 0.441 Qyzylorda Qyzylorda 0.002 0.003 0.166 0.020 0.881 0.017 9.558 0.360 Qyzylorda Shieliyskiy 0.000 0.000 0.164 0.039 0.889 0.031 8.916 0.486 Qyzylorda Terenozekskiy 0.000 0.000 0.160 0.040 0.962 0.020 7.819 0.299 37    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Qyzylorda Zhalagashskiy 0.000 0.000 0.100 0.041 0.984 0.016 7.784 0.314 Qyzylorda Zhanakorganskiy 0.000 0.000 0.126 0.032 0.956 0.020 7.889 0.325 South Kaz Arysskiy 0.000 0.000 0.000 0.000 0.817 0.029 11.340 0.325 South Kaz Baydibekskiy 0.000 0.000 0.016 0.012 0.866 0.032 10.730 0.353 South Kaz Chardarinskiy 0.000 0.000 0.311 0.045 1.000 0.000 7.146 0.224 South Kaz Kazygurtskiy 0.000 0.000 0.043 0.019 0.981 0.012 8.289 0.212 South Kaz Maktaaral`Skiy 0.000 0.000 0.056 0.014 0.968 0.011 8.212 0.171 South Kaz Ordabasynskiy 0.015 0.011 0.388 0.042 0.968 0.015 6.722 0.296 South Kaz Otrarskiy 0.000 0.000 0.000 0.000 0.781 0.041 12.435 0.501 South Kaz Saryagashskiy 0.014 0.006 0.332 0.024 0.981 0.007 6.859 0.144 South Kaz Sayramskiy 0.000 0.000 0.033 0.009 0.964 0.010 8.558 0.135 South Kaz Shymkent 0.000 0.000 0.016 0.005 0.783 0.015 11.774 0.151 South Kaz Suzakskiy * 0.000 0.085 0.000 0.147 1.014 0.079 9.984 1.813 South Kaz Tolebiyskiy 0.000 0.000 0.272 0.027 0.943 0.014 7.338 0.193 South Kaz Turkestan 0.000 0.000 0.061 0.010 0.950 0.009 8.447 0.106 South Kaz Tyul`Kubaskiy 0.000 0.000 0.132 0.034 0.926 0.025 8.841 0.340 West Kaz Akzhaikskiy 0.045 0.026 0.226 0.050 0.945 0.025 8.044 0.435 West Kaz Burlinskiy 0.025 0.018 0.091 0.033 0.823 0.040 10.478 0.552 West Kaz Chingirlauskiy 0.000 0.000 0.232 0.072 0.924 0.039 8.168 0.547 West Kaz Dzhangalinskiy * 0.009 0.081 0.143 0.141 0.897 0.075 9.563 1.794 West Kaz Dzhanybekskiy 0.000 0.000 0.209 0.065 0.953 0.029 7.578 0.468 West Kaz Karatobinskiy 0.000 0.000 0.104 0.052 0.880 0.048 9.280 0.643 West Kaz Kaztalovskiy 0.011 0.018 0.202 0.067 0.858 0.051 9.062 0.772 West Kaz Syrymskiy * 0.014 0.081 0.140 0.141 0.883 0.075 9.890 1.791 West Kaz Taskalinskiy 0.000 0.000 0.092 0.058 0.827 0.061 11.160 1.263 West Kaz Terektinskiy 0.034 0.025 0.253 0.059 0.909 0.034 8.028 0.651 West Kaz Urdinskiy 0.031 0.030 0.266 0.074 0.966 0.026 7.481 0.559 West Kaz Zelenovskiy 0.000 0.000 0.079 0.013 0.760 0.021 11.343 0.305 38    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Zhambyl Bayzakskiy 0.000 0.000 0.079 0.012 0.844 0.017 9.874 0.220 Zhambyl Kordayskiy 0.036 0.017 0.267 0.040 0.955 0.018 7.385 0.292 Zhambyl Lugovskoy 0.006 0.007 0.215 0.037 0.956 0.018 7.767 0.272 Zhambyl Merkenskiy 0.025 0.015 0.162 0.035 0.921 0.025 8.382 0.373 Zhambyl Moyynkumskiy 0.000 0.000 0.042 0.030 0.850 0.051 10.315 0.695 Zhambyl Sarysuskiy 0.029 0.021 0.268 0.052 0.973 0.020 6.744 0.340 Zhambyl Shuskiy 0.000 0.000 0.031 0.013 0.878 0.025 9.710 0.372 Zhambyl Talasskiy * 0.049 0.086 0.249 0.141 0.935 0.079 8.062 1.887 Zhambyl Zhamb. 0.033 0.016 0.443 0.042 0.994 0.007 6.188 0.194 Zhambyl Zhambylskiy * 0.062 0.086 0.231 0.142 0.905 0.079 8.108 1.885 Zhambyl Zhualy * 0.039 0.086 0.310 0.141 0.980 0.080 7.159 1.888 Zhambyl Zhualynskiy 0.000 0.000 0.032 0.016 0.956 0.018 8.663 0.261 39    Annex B – Uzbekistan District-level Results 2018 Table 3:Uzbekistan District-level Results 2018 Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Andijon Andijon 0.126 0.015 0.688 0.021 0.984 0.005 5.645 0.150 Andijon Asaka 0.160 0.014 0.559 0.019 0.960 0.007 6.290 0.150 Andijon Baliqchi 0.127 0.015 0.689 0.021 0.945 0.011 5.673 0.159 Andijon Bo'Zsuv * 0.105 0.087 0.640 0.150 0.935 0.078 6.731 1.837 Andijon Buloqboshi * 0.140 0.087 0.566 0.144 0.906 0.076 7.109 1.813 Andijon Izboskan * 0.108 0.087 0.523 0.146 0.912 0.077 7.099 1.834 Andijon Jalolquduq 0.095 0.011 0.559 0.019 0.902 0.011 7.192 0.218 Andijon Marhamat * 0.013 0.091 0.368 0.142 0.858 0.076 8.045 1.797 Andijon Oltinko'L * 0.129 0.087 0.692 0.157 0.935 0.078 6.664 1.861 Andijon Paxtaobod 0.057 0.011 0.482 0.023 0.949 0.010 8.387 0.475 Andijon Qo'Rg'Ontepa * 0.032 0.091 0.385 0.142 0.834 0.076 8.602 1.804 Andijon Shahrixon 0.175 0.018 0.645 0.022 0.882 0.015 7.724 0.456 Andijon Ulug'Nor * 0.016 0.090 0.455 0.145 0.890 0.076 7.622 1.798 Andijon Xo'Jaobod 0.048 0.014 0.513 0.032 0.942 0.015 7.389 0.287 Bukhoro Buxoro 0.065 0.009 0.328 0.017 0.872 0.012 8.816 0.252 Bukhoro G'Ijduvon * 0.126 0.082 0.297 0.142 0.859 0.081 10.581 1.942 Bukhoro Jondor * 0.085 0.082 0.275 0.142 0.829 0.081 10.953 1.941 Bukhoro Kogon 0.055 0.014 0.318 0.030 0.817 0.024 10.174 0.630 Bukhoro Olot 0.024 0.013 0.162 0.032 0.737 0.037 11.651 1.565 Bukhoro Peshku 0.043 0.013 0.158 0.023 0.773 0.026 13.428 0.842 Bukhoro Qorako'L * 0.088 0.082 0.255 0.142 0.832 0.081 10.949 1.943 Bukhoro Qorovulbozor * 0.117 0.082 0.304 0.142 0.830 0.081 10.592 1.940 Bukhoro Romitan 0.158 0.032 0.496 0.043 0.800 0.033 10.814 1.028 Bukhoro Shofirkon 0.106 0.019 0.342 0.030 0.828 0.023 11.649 0.808 Bukhoro Vobkent 0.110 0.014 0.345 0.021 0.851 0.016 9.248 0.367 40    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Ferghana Beshariq 0.231 0.026 0.574 0.031 0.905 0.018 7.450 0.373 Ferghana Bog'Dod 0.107 0.010 0.431 0.016 0.882 0.010 7.923 0.175 Ferghana Buvayda 0.092 0.018 0.439 0.032 0.883 0.020 8.144 0.551 Ferghana Dang'Ara 0.064 0.009 0.293 0.018 0.884 0.012 8.895 0.229 Ferghana Farg'Ona 0.066 0.010 0.304 0.018 0.858 0.013 9.612 0.274 Ferghana Furqat * 0.124 0.082 0.337 0.142 0.865 0.076 8.002 1.800 Ferghana O'Zbekiston 0.048 0.008 0.303 0.017 0.867 0.013 8.558 0.178 Ferghana Oltiariq 0.134 0.021 0.351 0.030 0.912 0.018 8.314 0.302 Ferghana Oxunboboev 0.000 0.000 0.281 0.028 0.660 0.029 10.205 0.368 Ferghana Quva 0.004 0.004 0.170 0.024 0.693 0.028 11.834 0.448 Ferghana Rishton * 0.145 0.082 0.353 0.142 0.873 0.075 7.945 1.798 Ferghana So’X * 0.076 0.082 0.234 0.142 0.889 0.075 9.312 1.825 Ferghana Toshloq * 0.218 0.084 0.584 0.149 0.902 0.077 7.211 1.821 Ferghana Uchko'Prik * 0.121 0.082 0.436 0.151 0.848 0.076 8.333 1.798 Ferghana Yozyovon * 0.139 0.082 0.485 0.150 0.867 0.076 7.788 1.798 Jizzakh Arnasoy * 0.009 0.089 0.254 0.154 0.796 0.083 10.417 1.964 Jizzakh Baxmal * 0.093 0.089 0.353 0.155 0.772 0.083 9.719 1.976 Jizzakh Do'Stlik * 0.009 0.089 0.253 0.154 0.797 0.083 10.406 1.964 Jizzakh Forish 0.000 0.000 0.324 0.033 0.768 0.029 11.639 0.667 Jizzakh G'Allaorol * 0.033 0.089 0.257 0.154 0.785 0.083 10.616 1.966 Jizzakh Jizzax 0.045 0.009 0.152 0.015 0.756 0.018 11.013 0.264 Jizzakh Mirzacho'L * 0.010 0.089 0.254 0.154 0.806 0.083 10.230 1.964 Jizzakh Paxtakor 0.030 0.012 0.211 0.029 0.870 0.023 9.430 0.330 Jizzakh Yangiobod * 0.000 0.089 0.233 0.154 0.812 0.083 10.510 1.966 Jizzakh Zafarobod * 0.051 0.089 0.273 0.154 0.799 0.083 10.209 1.965 Jizzakh Zarbdor 0.000 0.000 0.149 0.025 0.776 0.029 9.280 0.273 Jizzakh Zomin 0.051 0.016 0.235 0.030 0.729 0.030 12.476 0.755 Karakalpakstan Amudaryo 0.376 0.043 0.689 0.044 0.952 0.021 5.839 0.533 Karakalpakstan Beruniy 0.135 0.032 0.586 0.047 0.952 0.021 6.176 0.317 41    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Karakalpakstan Chimboy 0.273 0.025 0.494 0.029 0.947 0.013 7.196 0.432 Karakalpakstan Ellikqala * 0.099 0.087 0.460 0.151 0.946 0.081 7.467 1.924 Karakalpakstan Kegeyli 0.104 0.021 0.494 0.035 0.981 0.010 6.741 0.299 Karakalpakstan Mo'Ynoq * 0.001 0.091 0.334 0.158 0.880 0.086 8.739 2.038 Karakalpakstan Nukus * 0.185 0.086 0.555 0.149 0.957 0.079 6.032 1.885 Karakalpakstan Qanliko'L * 0.180 0.086 0.568 0.149 0.958 0.079 6.009 1.882 Karakalpakstan Qo'Ng'Irot 0.055 0.021 0.434 0.047 1.000 0.000 5.981 0.224 Karakalpakstan Qorao'Zak * 0.147 0.086 0.517 0.149 0.968 0.079 6.713 1.886 Karakalpakstan Shumanay 0.146 0.032 0.415 0.047 1.000 0.000 6.314 0.263 Karakalpakstan Taxtako'Pir 0.117 0.031 0.682 0.044 1.000 0.000 5.152 0.219 Karakalpakstan To'Rtko'L 0.085 0.014 0.453 0.025 0.953 0.011 7.085 0.244 Karakalpakstan Xo'Jayli 0.192 0.016 0.517 0.021 0.962 0.008 6.252 0.147 Kashkadarya Chiroqchi * 0.096 0.085 0.369 0.142 0.848 0.075 8.831 1.802 Kashkadarya Dehqonobod 0.000 0.000 0.257 0.031 0.865 0.023 9.192 0.449 Kashkadarya G'Uzor 0.174 0.027 0.371 0.034 0.895 0.021 7.487 0.423 Kashkadarya Kasbi 0.109 0.013 0.329 0.019 0.800 0.016 8.608 0.204 Kashkadarya Kitob * 0.081 0.085 0.385 0.142 0.879 0.075 7.893 1.799 Kashkadarya Koson 0.032 0.009 0.371 0.024 0.830 0.019 10.298 0.498 Kashkadarya Muborak * 0.030 0.085 0.295 0.141 0.887 0.076 8.802 1.795 Kashkadarya Nishon 0.032 0.009 0.286 0.023 0.771 0.021 12.201 0.622 Kashkadarya Qamashi 0.055 0.010 0.241 0.018 0.772 0.017 11.096 0.349 Kashkadarya Qarshi 0.008 0.004 0.114 0.013 0.726 0.017 11.498 0.279 Kashkadarya Shahrisabz 0.087 0.020 0.382 0.034 0.960 0.013 7.575 0.288 Kashkadarya Usmon Yusupov 0.144 0.025 0.309 0.033 0.778 0.029 9.466 0.381 Kashkadarya Yakkabog' 0.103 0.016 0.358 0.025 0.773 0.021 10.554 0.483 Khorezm Bog'Ot * 0.059 0.087 0.385 0.150 0.893 0.075 8.587 1.796 Khorezm Gurlan 0.110 0.018 0.417 0.028 0.965 0.010 6.795 0.195 Khorezm Hazorasp 0.030 0.008 0.341 0.022 0.972 0.008 7.053 0.146 Khorezm Qo'Shko'Pir 0.086 0.022 0.309 0.036 0.727 0.034 12.658 0.959 42    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Khorezm Shovot 0.046 0.012 0.475 0.028 0.944 0.013 6.463 0.203 Khorezm Urganch 0.075 0.009 0.361 0.017 0.885 0.011 9.767 0.366 Khorezm Xiva 0.000 0.000 0.353 0.040 0.921 0.022 7.088 0.264 Khorezm Xonqa * 0.066 0.087 0.397 0.150 0.897 0.076 8.468 1.797 Khorezm Yangiariq * 0.065 0.087 0.395 0.150 0.894 0.076 8.492 1.796 Khorezm Yangibozor 0.091 0.023 0.448 0.039 0.929 0.020 7.587 0.585 Namangan Chortoq 0.076 0.011 0.213 0.017 0.777 0.017 12.201 0.598 Namangan Chust 0.093 0.015 0.372 0.025 0.839 0.018 9.445 0.426 Namangan Kosonsoy 0.006 0.005 0.184 0.027 0.811 0.027 9.361 0.313 Namangan Mingbuloq 0.157 0.025 0.473 0.035 0.909 0.020 7.183 0.356 Namangan Namangan 0.191 0.014 0.434 0.017 0.872 0.012 8.271 0.252 Namangan Norin * 0.273 0.091 0.647 0.150 0.949 0.080 6.457 1.887 Namangan Pop 0.113 0.022 0.255 0.031 0.838 0.025 12.802 0.963 Namangan To'Raqo'Rg'On * 0.193 0.084 0.500 0.145 0.904 0.076 7.754 1.814 Namangan Uchqo'Rg'On 0.000 0.000 0.251 0.031 0.890 0.022 9.144 0.353 Namangan Uychi 0.124 0.023 0.322 0.033 0.873 0.023 9.521 0.797 Namangan Yagiqo'Rg'On 0.114 0.016 0.376 0.025 0.868 0.017 7.673 0.221 Navoi Karmana * 0.149 0.082 0.309 0.142 0.907 0.076 8.486 1.791 Navoi Konimex * 0.115 0.082 0.269 0.142 0.939 0.077 8.726 1.792 Navoi Navbahor * 0.096 0.082 0.265 0.142 0.840 0.076 8.850 1.799 Navoi Nurota 0.030 0.007 0.300 0.019 0.857 0.015 9.306 0.335 Navoi Qiziltepa 0.029 0.011 0.276 0.029 0.754 0.028 10.507 1.606 Navoi Tomdi 0.094 0.018 0.352 0.030 0.924 0.016 7.939 0.449 Navoi Uchquduq * 0.000 0.083 0.094 0.144 0.916 0.077 9.760 1.840 Navoi Xatirchi 0.138 0.018 0.486 0.026 0.918 0.014 7.579 0.471 Samarkand Bulung'Ur * 0.181 0.082 0.424 0.142 0.826 0.076 8.012 1.804 Samarkand Ishtixon 0.144 0.011 0.374 0.015 0.848 0.011 10.228 0.311 Samarkand Jomboy * 0.258 0.084 0.558 0.144 0.883 0.076 6.894 1.819 Samarkand Kattaqo'Rg'On 0.230 0.023 0.483 0.028 0.915 0.015 8.686 0.696 43    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Samarkand Narpay * 0.190 0.082 0.440 0.142 0.900 0.075 7.998 1.797 Samarkand Nurobod 0.352 0.028 0.716 0.027 1.000 0.000 4.643 0.140 Samarkand Oqdaryo 0.246 0.024 0.348 0.027 0.794 0.023 10.520 0.547 Samarkand Pastdarg'Om 0.080 0.012 0.226 0.018 0.744 0.019 10.954 0.416 Samarkand Paxtachi * 0.147 0.082 0.343 0.142 0.907 0.076 8.550 1.792 Samarkand Payariq * 0.213 0.082 0.469 0.143 0.869 0.075 7.874 1.806 Samarkand Qo'Shrabot * 0.182 0.082 0.414 0.142 0.877 0.075 8.291 1.792 Samarkand Samarqand 0.098 0.018 0.450 0.030 0.906 0.017 7.208 0.250 Samarkand Toyloq 0.165 0.010 0.470 0.014 0.901 0.008 7.525 0.185 Samarkand Urgut * 0.200 0.082 0.387 0.142 0.850 0.075 8.469 1.797 Sirdaryo Boyovut 0.134 0.024 0.417 0.035 0.847 0.025 7.723 0.388 Sirdaryo Guliston * 0.124 0.082 0.348 0.142 0.871 0.076 8.579 1.794 Sirdaryo Mirzaobod 0.087 0.020 0.317 0.033 0.944 0.016 7.457 0.344 Sirdaryo Oqoltin 0.117 0.016 0.464 0.025 0.980 0.007 6.899 0.190 Sirdaryo Sayxunobod * 0.141 0.082 0.361 0.142 0.887 0.075 8.172 1.795 Sirdaryo Sharof Rashidov * 0.125 0.082 0.360 0.142 0.865 0.076 8.553 1.793 Sirdaryo Sirdaryo 0.162 0.026 0.498 0.035 0.955 0.014 6.361 0.264 Sirdaryo Xovos 0.134 0.024 0.623 0.034 0.967 0.012 6.071 0.264 Surkhandarya Angor 0.079 0.015 0.278 0.026 0.926 0.015 7.286 0.213 Surkhandarya Bandixon * 0.240 0.085 0.688 0.152 0.927 0.076 6.827 1.820 Surkhandarya Boysun * 0.162 0.082 0.464 0.150 0.858 0.075 9.433 1.796 Surkhandarya Denov 0.260 0.024 0.488 0.028 0.915 0.015 6.344 0.254 Surkhandarya Jarqo'Rg'On * 0.168 0.082 0.533 0.149 0.890 0.075 7.980 1.798 Surkhandarya Muzrabot 0.154 0.014 0.427 0.020 0.792 0.016 9.724 0.351 Surkhandarya Oltinsoy * 0.171 0.082 0.542 0.150 0.845 0.075 8.315 1.805 Surkhandarya Qiziriq * 0.178 0.082 0.579 0.149 0.906 0.075 7.624 1.804 Surkhandarya Qumqo'Rg'On 0.281 0.025 0.371 0.027 0.821 0.021 8.914 0.472 Surkhandarya Sariosiyo 0.077 0.015 0.474 0.028 0.947 0.012 6.193 0.159 Surkhandarya Sherobod 0.107 0.017 0.743 0.025 0.963 0.011 5.862 0.196 44    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Surkhandarya Sho'Rchi * 0.159 0.082 0.531 0.149 0.891 0.075 8.372 1.796 Surkhandarya Termiz 0.162 0.017 0.605 0.022 0.904 0.013 9.731 1.367 Surkhandarya Uzun 0.120 0.018 0.480 0.028 0.997 0.003 5.912 0.152 Tashkent Bekobod 0.102 0.020 0.288 0.029 0.798 0.025 8.585 0.304 Tashkent Bo'Ka * 0.062 0.087 0.343 0.142 0.857 0.076 8.602 1.793 Tashkent Bo'Stonliq 0.020 0.009 0.034 0.012 0.610 0.030 14.101 0.384 Tashkent Chinoz * 0.138 0.087 0.435 0.144 0.891 0.076 7.830 1.812 Tashkent O'Rtachirchiq 0.000 0.000 0.209 0.026 0.699 0.029 10.584 0.354 Tashkent Ohangaron 0.079 0.013 0.228 0.021 0.914 0.013 8.870 0.211 Tashkent Oqqo'Rg'On 0.000 0.000 0.254 0.028 0.883 0.020 8.595 0.365 Tashkent Parkent * 0.045 0.087 0.288 0.142 0.791 0.076 10.377 1.809 Tashkent Piskent 0.050 0.008 0.310 0.018 0.758 0.016 12.693 0.622 Tashkent Qibray 0.025 0.004 0.144 0.009 0.697 0.012 11.555 0.187 Tashkent Quyichirchiq * 0.137 0.087 0.471 0.143 0.897 0.076 7.801 1.815 Tashkent Toshkent 0.390 0.031 0.832 0.024 0.979 0.009 4.871 0.297 Tashkent Yangiyo'L * 0.129 0.087 0.427 0.143 0.885 0.076 7.975 1.808 Tashkent Yuqorichirchiq * 0.068 0.087 0.364 0.142 0.820 0.076 9.076 1.802 Tashkent Zangiota * 0.162 0.088 0.487 0.146 0.887 0.077 7.549 1.833 Tashkent City Tashkent City 0.000 0.000 0.165 0.007 0.736 0.009 11.953 0.172 45    Annex C – Kyrgyz Republic District-level Results 2016 Table 4: Kyrgyz Republic District-level Results 2016 Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Batken Batken 0.089 0.025 0.669 0.044 0.982 0.012 5.664 0.376 Batken Kadamjai 0.297 0.029 0.785 0.027 0.995 0.005 4.385 0.150 Batken Lailak 0.310 0.033 0.787 0.030 0.994 0.006 4.388 0.159 BišKek BišKek 0.165 0.011 0.516 0.015 0.958 0.006 6.392 0.110 ChüY AlamüDüN 0.102 0.019 0.567 0.032 0.980 0.009 5.712 0.183 ChüY Chui 0.279 0.040 0.725 0.043 0.995 0.007 4.607 0.227 ChüY Jaiyl 0.069 0.021 0.367 0.041 0.940 0.021 7.410 0.394 ChüY Kemin 0.136 0.037 0.823 0.042 0.999 0.004 4.435 0.169 ChüY Moskovsky 0.186 0.035 0.765 0.038 0.974 0.015 4.807 0.283 ChüY Panfilov 0.000 0.000 0.237 0.061 0.968 0.032 7.705 0.651 ChüY Sokuluk 0.270 0.028 0.630 0.031 0.987 0.007 5.156 0.194 ChüY Ysyk-Ata 0.245 0.034 0.641 0.039 0.987 0.009 5.147 0.242 Jalal-Abad Aksyi 0.301 0.030 0.833 0.025 1.000 0.001 4.156 0.107 Jalal-Abad Ala-Buka 0.048 0.018 0.676 0.042 0.992 0.009 5.299 0.189 Jalal-Abad Bazar-Korgon 0.164 0.027 0.599 0.037 0.995 0.005 5.138 0.143 Jalal-Abad Chatkal 0.079 0.024 0.593 0.044 0.997 0.005 5.581 0.201 Jalal-Abad Nooken 0.229 0.031 0.729 0.033 0.986 0.009 5.027 0.396 Jalal-Abad Suzak 0.226 0.019 0.781 0.019 0.993 0.004 4.614 0.105 Jalal-Abad Togus-Toro * . . . . . . . . Jalal-Abad Toktogul 0.141 0.029 0.716 0.039 0.996 0.005 4.809 0.166 Naryn Ak-Talaa 0.421 0.054 0.912 0.033 0.997 0.006 3.253 0.229 Naryn At-Bashi 0.107 0.037 0.481 0.067 0.941 0.028 6.868 0.768 Naryn Jumgal 0.150 0.038 0.795 0.042 0.995 0.007 4.560 0.244 Naryn Kochkor 0.313 0.047 0.718 0.049 0.952 0.021 5.166 0.442 Naryn Naryn 0.210 0.038 0.624 0.047 0.971 0.015 5.558 0.361 Osh Alai 0.036 0.015 0.543 0.042 0.975 0.014 6.236 0.253 46    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Osh Aravan 0.333 0.034 0.837 0.028 0.995 0.006 4.189 0.156 Osh Chong-Alay * . . . . . . . . Osh Kara-Kuldja 0.093 0.026 0.775 0.040 0.973 0.016 4.941 0.275 Osh Kara-Suu 0.174 0.017 0.745 0.020 0.986 0.005 4.833 0.106 Osh Nookat 0.064 0.014 0.716 0.025 0.991 0.005 5.264 0.125 Osh Uzgen 0.278 0.025 0.812 0.022 0.996 0.003 4.331 0.117 Osh (City) Osh 0.171 0.021 0.704 0.025 0.978 0.008 5.318 0.176 Talas Bakai-Ata 0.018 0.016 0.351 0.057 0.991 0.013 7.025 0.322 Talas Kara-Buura 0.231 0.043 0.826 0.039 0.995 0.007 4.208 0.195 Talas Manas 0.012 0.015 0.583 0.070 0.996 0.011 5.645 0.335 Talas Talas 0.053 0.019 0.668 0.042 0.994 0.007 5.379 0.226 Ysyk-KöL Ak-Suu 0.090 0.021 0.499 0.038 0.953 0.016 6.678 0.310 Ysyk-KöL Djety-Oguz 0.047 0.019 0.582 0.047 0.970 0.017 6.135 0.338 Ysyk-KöL Ton 0.350 0.049 0.884 0.036 0.999 0.005 3.732 0.185 Ysyk-KöL TüP 0.296 0.050 0.825 0.047 0.997 0.006 4.297 0.242 Ysyk-KöL Ysyk-KöL 0.170 0.028 0.632 0.037 0.973 0.013 5.587 0.289 47    Annex D – Tajikistan District-level Results 2015 Table 5:Tajikistan District-level Results 2015 Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Dushanbe Rudaki 0.053 0.008 0.305 0.016 0.841 0.013 9.121 0.212 GBAO Darvoz * 0.188 0.082 0.369 0.142 0.816 0.076 9.096 1.808 GBAO Ishkoshim * 0.296 0.085 0.577 0.152 0.923 0.077 6.435 1.960 GBAO Murghob * 0.245 0.091 0.632 0.161 0.835 0.080 7.530 2.052 GBAO Roshtqala * 0.294 0.085 0.575 0.151 0.927 0.077 6.455 1.958 GBAO Rushon 0.246 0.034 0.707 0.036 1.000 0.000 4.641 0.166 GBAO Shughnon 0.199 0.035 0.685 0.044 0.989 0.010 5.470 0.284 GBAO Vanj * 0.153 0.083 0.348 0.145 0.854 0.080 10.010 1.849 Khatlon Baljuvon 0.070 0.028 0.347 0.055 0.935 0.030 7.610 0.461 Khatlon Bokhtar 0.339 0.020 0.702 0.020 0.996 0.003 4.644 0.108 Khatlon Danghara * 0.199 0.082 0.592 0.144 0.954 0.077 6.322 1.837 Khatlon Farkhor 0.174 0.027 0.609 0.036 0.971 0.013 5.599 0.238 Khatlon Jilikul 0.375 0.051 0.956 0.023 1.000 0.000 3.330 0.150 Khatlon Jomi 0.327 0.033 0.641 0.035 0.995 0.005 4.777 0.243 Khatlon Khovaling 0.000 0.000 0.253 0.037 0.845 0.032 9.152 0.745 Khatlon Khuroson 0.258 0.037 0.633 0.043 1.000 0.000 5.130 0.272 Khatlon Kolkhozobod 0.137 0.029 0.593 0.043 0.941 0.022 6.325 0.392 Khatlon Kulob 0.375 0.025 0.728 0.023 0.979 0.007 4.687 0.177 Khatlon Moskva * 0.168 0.082 0.552 0.144 0.955 0.077 5.965 1.831 Khatlon Muminobod * . . . . . . . . Khatlon Norak 0.048 0.017 0.236 0.036 0.892 0.028 8.091 0.378 Khatlon Nosir Khusrav 0.212 0.046 0.625 0.058 0.986 0.016 5.954 0.440 Khatlon Panj 0.058 0.016 0.504 0.037 0.963 0.014 6.452 0.250 Khatlon Qabodiyon 0.000 0.000 0.358 0.056 1.000 0.000 7.107 0.309 Khatlon Qumsangir 0.335 0.029 0.763 0.027 0.974 0.010 4.658 0.236 48    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. Khatlon Sarband * . . . . . . . . Khatlon Shahrituz 0.089 0.020 0.473 0.036 0.928 0.019 6.832 0.298 Khatlon Shurobod 0.222 0.046 0.770 0.049 1.000 0.000 4.249 0.205 Khatlon Sovet 0.434 0.034 0.793 0.030 1.000 0.000 4.062 0.162 Khatlon Vakhsh * . . . . . . . . Khatlon Vose 0.406 0.034 0.768 0.031 0.996 0.005 4.042 0.144 Khatlon Yovon 0.298 0.032 0.634 0.035 0.987 0.008 5.071 0.212 Sogd Asht 0.132 0.025 0.474 0.036 0.940 0.017 7.743 0.573 Sogd Ayni 0.000 0.000 0.285 0.052 0.824 0.045 9.771 0.951 Sogd Ghafurov 0.111 0.010 0.451 0.017 0.896 0.010 7.423 0.171 Sogd Ghonchi * 0.049 0.082 0.404 0.148 0.933 0.078 7.912 1.875 Sogd Isfara 0.080 0.023 0.472 0.044 0.953 0.019 6.861 0.403 Sogd Istaravshan 0.328 0.020 0.724 0.019 0.977 0.006 4.635 0.125 Sogd Jabor Rasulov 0.211 0.048 0.852 0.040 1.000 0.000 4.664 0.328 Sogd Konibodom 0.167 0.032 0.547 0.044 0.964 0.017 6.117 0.301 Kuhistoni Sogd Mastchoh 0.184 0.041 0.311 0.052 0.800 0.047 9.967 0.903 Sogd Matchin 0.029 0.014 0.605 0.043 1.000 0.000 5.359 0.176 Sogd Pandjakent 0.250 0.024 0.617 0.028 0.953 0.012 5.547 0.228 Sogd Shahriston 0.101 0.036 0.641 0.057 0.971 0.020 5.768 0.412 Sogd Spitamen 0.228 0.048 0.747 0.051 1.000 0.000 4.552 0.249 Sogd Zafarobod 0.266 0.049 0.801 0.047 1.000 0.000 3.762 0.178 RRS Fayzobod 0.049 0.015 0.450 0.037 0.974 0.012 6.481 0.209 RRS Hissor 0.149 0.020 0.444 0.029 0.922 0.016 6.862 0.258 RRS Jirgatol 0.000 0.000 0.197 0.035 0.833 0.034 9.882 0.483 RRS Nurobod 0.707 0.034 0.933 0.022 0.997 0.005 2.725 0.157 RRS Rasht 0.431 0.031 0.666 0.030 0.982 0.009 4.697 0.243 RRS Roghun * . . . . . . . . RRS Rudaki 0.153 0.023 0.522 0.032 0.961 0.013 6.581 0.271 49    Not Rate Rate Below Mean Province District Precise ($3.2) S.E. ($5.5) S.E. MClass S.E. Cons. S.E. RRS Shahrinav 0.114 0.037 0.596 0.060 0.973 0.020 5.858 0.443 RRS Tavildara * 0.148 0.083 0.508 0.149 0.960 0.080 6.925 1.890 RRS Tojikobod * . . . . . . . . RRS Tursunzoda 0.140 0.022 0.594 0.032 0.982 0.009 5.430 0.172 RRS Vahdat 0.036 0.008 0.241 0.019 0.925 0.011 8.299 0.192 RRS Varzob 0.115 0.038 0.423 0.060 1.000 0.000 6.525 0.434 50    Annex E – Country-Level Maps Figure 17: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Kazakhstan 2017 51    Figure 18: Share of the population in the middle class, rates for Kazakhstan 2017 52    Figure 19: Average Per Capita Daily Consumption in 2011 PPP for Kazakhstan 2017 53    Figure 20: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Uzbekistan 2018 Figure 21: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Uzbekistan 2018 55    Figure 22: Share of population in the middle class, rates for Uzbekistan 2018 56    Figure 23: Average Per Capita Daily Consumption in 2011 PPP for Uzbekistan 2018 57    Figure 24: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Kyrgyzstan 2016 58    Figure 25: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Kyrgyzstan 2016 59    Figure 26: Share of the population in the middle class, rates for Kyrgyzstan 2016 60    Figure 27: Average Per Capita Daily Consumption in 2011 PPP for Kyrgyzstan 2016 61    Figure 28: Poverty headcount ratio at $3.2-a-day PPP per person, rates for Tajikistan 2015 62    Figure 29: Poverty headcount ratio at $5.5-a-day PPP per person, rates for Tajikistan 2015 63    Figure 30: Share of the population in the middle class, rates for Tajikistan 2015 64    Figure 31: Average Per Capita Daily Consumption in 2011 PPP for Tajikistan 2015 65    Annex F – Validation and Precision One concern often raised regarding poverty map results is that the poverty estimates may not be precise enough for the purposes to which they are put. One way to assess this is to measure whether there are clear and statistically significant differences between the resulting poverty rates at the district level that can be usefully integrated into the work of the World Bank, counterparts, and other partners. This can be demonstrated graphically by reporting the poverty rates with the associated 95% confidence interval. Because there are some districts in which there are no survey observations, these must be imputed out of sample when using the FH approach. In this minority of cases, the root mean square errors are larger than average, and should be used with caution. These cases of relatively high imprecision are noted in the district level tables with red stars (*). However, it is also important to note that most cases in this application are in sample do not suffer from this limitation of the FH approach. Figure 32: Upper and Lower-bound Confidence Intervals, Including Out-of-Sample Predictions 1 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 fh_55 upper_kaz lower_kaz upper_kgz lower_kgz upper_tjk lower_tjk upper_uzb lower_uzb Another way to demonstrate performance is to benchmark against “gold standard” results using the Elbers, Lanjouw, and Lanjouw (ELL) method previously published by the World Bank. The following graph demonstrates the comparison visually for 17 previously published maps that used the ELL approach. The results from the Central Asia poverty map perform quite well in this comparison for those cases that are in-sample. The 95% confidence intervals are usually well within those of other published results. However, there are notable exceptions for cases in which the estimates are derived entirely from the FH model (i.e. those districts in which there were no observations in the underlying survey). The root mean square errors for these cases are higher than in the case for most ELL maps and this is one of the key advantages of the ELL approach (because such analyses use census microdata, there are no missing districts that need to be estimated out-of-sample). However, unfortunately, in this application the ELL approach is not possible. Figure 33: Comparison between ELL and FH 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% ELL Maps Lower ELL Maps Upper CA FH Lower CA FH Upper The improvements in precision (as measured by root mean square error) over direct survey estimates from the Fay Herriot approach are also substantial. Improvements are concentrated in those locations that had the least precise estimates before the exercise. The following figure (34) illustrates the improvement in terms of the root mean square error of the mean for poverty at $5.5/day. 67    Figure 34: Root Mean Square Error Before and After FH Approach (Poverty Headcount Ratio $5.5/day - Left; Average per-capita consumption 2011 PPP – Right) 0.12 4 3.5 0.1 3 0.08 2.5 0.06 2 1.5 0.04 1 0.02 0.5 0 0 Standard Error Fay‐Herriot Standard Error Fay‐Herriot Standard Error No Correction Standard Error No Correction An additional robustness check estimates FH models for each country individually, comparing the results with those using pooled data across countries. The results presented in figure (35) show that this makes little substantive difference in either the point estimates or in the root mean square error for within-sample districts. There is a very strong relationship between the estimates regardless of the pooled vs. individual approach. The relationship remains strong but weakens somewhat with respect to out-of-sample districts (figure 35 - right). Using the pooled approach likely improves on the out- of-sample prediction: this method improves the training of the model with more data than is available at the national level. Figure 35: Comparison of Individual Models vs. Pooled Model (Only in sample - Left; Including out of sample - Right) 1 1 0.9 R² = 0.9852 0.9 R² = 0.9216 0.8 0.8 0.7 0.7 0.6 0.6 Pooled Pooled 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Individual Country Maps Individual Country Maps 68    These results suggest that there is little concern that country-specific idiosyncratic relationships between y and the explanatory variables are contaminating the results for other countries. Because the right-hand side variables in the model are almost entirely based on remote-sensing data, and are not country-specific in nature, the approach is more amenable for such cross-country modeling than other potential methods that are more sensitive to these concerns. None of the explanatory variables that are used in the modeling stage (reviewed in annex H) in the final regression models would be expected to vary directly with policy regimes in the countries under study. The model selection adopted in this approach uses the stepwise approach, which is vulnerable to the inclusion of variables with high variance inflation factors. Comparing a model that screens for high VIFs is reported in figure (36). The results show that limiting the variance inflation factor (VIF) of the variables included in the stepwise selection approach to standard levels has some small effect on the model and consequently the estimated results, and a reduction in the adjusted R2 (for the $5.5/day line, down from .72 to .67, when estimated over all Central Asia). Though this change leads to only small differences in terms of the point estimates or the root mean square error of the estimation results, the VIF minimizing strategy is clearly superior, and adopted in this report. Figure 36: Left: Strict VIF Threshold (5) vs. Adopted Threshold (10); Right: No VIF Threshold vs. Adopted Threshold (1)) 1 1 R² = 0.9758 R² = 0.9628 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 To ensure that the map faithfully represents the underlying poverty dynamics of the country, it is important to ensure that the results are internally consistent. One method of assessing the stability of the estimates and the robustness of the results is to compare the relationship between the direct and adjusted estimates. This relationshipcan be presented visually as a scatter plot, with direct survey estimates along the x-axis, and model estimates along the y-axis. With perfect correlation, the estimates would lie along the 45˚ line, and as figure (37) shows, there is indeed a strong relationship between the model predictions and those of the observed values in the survey data. The cases in which the FH most differs from the direct estimate, the direct estimate was relatively imprecise, and the synthetic result thus relies more on the model-based estimate of the indicator. The overall strong correspondence suggests that the model-based estimates improve the precision of estimates without 69    dramatically changing the point estimates derived from the survey. In some cases, the estimate lies exactly on the 45˚ which signifies that the model-based estimates did not provide additional information enough to revise the direct estimates. Figure 37: Direct vs. Fay Herriot Estimates for in-sample Poverty $5.5day (left) and average consumption (right) 1 20 R² = 0.9975 R² = 0.9628 0.9 18 0.8 16 0.7 14 Fay Herriot 12 Fay Herriot 0.6 0.5 10 0.4 8 0.3 6 0.2 4 0.1 2 0 0 0 0.2 0.4 0.6 0.8 1 0 5 10 15 20 Direct Estimate Direct Estimate Coefficients of variation (CVs) are a useful guideline with respect to precision when estimates are far away from zero, and when the sample size is small (for instance, the surveys in UZB and TJK are much smaller than for the other two countries). In the case of per-capita consumption, no resulting in-sample estimates are above the 20% threshold that some statistical agencies adopt for reporting (Eurostat, 2013). Furthermore, the R2 between district poverty rates at $5.5/day and average consumption is very high (.74) suggesting that the estimates of poverty and average consumption are very consistent. This strong accordance should give greater credence to the poverty estimates. 70    Figure 38: Improvement of Coefficient of Variation in average consumption 18 25 16 14 20 12 15 10 8 10 6 4 5 2 0 0 Avg Consumption Day/PPP (Left) Fay Herriot CV (%) Direct CV (%) Linear (Fay Herriot CV (%)) Linear (Direct CV (%)) 71    Annex G – Comparison of Domain Variance Estimation Alternatives The preceding results use the linearized variance estimator approach—based on a first-order Taylor series. The two other options are contrasted with this approach in figures (39) and (40) for average consumption and the poverty rate at $5.5/day respectively. Figure (41) provides comparisons of poverty rate results comparing across methods of variance estimation, highlighting the relatively small differences between the adopted approach (though the largest differences are present for the HT- based approach). Figure 39 comparison of variance estimates for average consumption 100 100 90 90 80 80 70 70 60 60 50 50 40 40 30 30 20 20 10 10 0 0 0 100 200 300 0 100 200 300 HT Variance Est. Taylor Linearization Smoothed WB Taylor Linearization Figure 40: Comparison of variance estimates for poverty at $5.5/day 0.08 0.08 0.07 0.07 0.06 0.06 0.05 0.05 0.04 0.04 0.03 0.03 0.02 0.02 0.01 0.01 0 0 0 100 200 300 0 100 200 300 Taylor Linearization HT Variance Est. Taylor Linearization Smoothed WB 72    Figure 41: Poverty Headcount Ratio (5.5/day) for Taylor Linearized Method vs. HT Method (Left), Taylor Liberalized Method vs. Smoothed (Right) 100% 100% R² = 0.9083 R² = 0.9909 90% 90% 80% 80% 70% 70% 60% 60% 50% 50% 40% 40% 30% 30% 20% 20% 10% 10% 0% 0% 0% 20% 40% 60% 80% 100% 0% 20% 40% 60% 80% 100% 73    Annex H – Regression Model Details Table 6: Fay-Herriot Model for Poverty Rate at $5.5/day and $3.2/day Poor Poor $5.5/day $3.2/day (1) (2) No local service road max value 1.0000*** No local service road max value 1.0000*** (0.0000) (0.0000) Grid Mean GDP 0.9559*** Minimum air temperature 1.0088*** (0.0103) (0.0012) Grid Mean Max 0.9981*** Maximum elevation slope 1.0024*** (0.0005) (0.0006) Air temperature max 1.0115*** Grid minimum GDP 1.4635*** (0.0017) (0.1916) Grid Minimum GDP 1.7417*** 500m Elevation 1.0001*** (0.3025) (0.0000) Minimum population count 1.0002** Grid mean GDP 0.9702*** (0.0001) (0.0083) 500m Elevation 1.0003*** Constant 1.0155 (0.0001) (0.0155) Constant 1.1588*** (0.0373) Observations 331 Observations 331 R-squared 0.6984 R-squared 0.4538 *** p<0.01, ** p<0.05, * p<0.1; selected *** p<0.01, ** p<0.05, * p<0.1; selected country/region dummy variables not shown country/region dummy variables not shown Table 7: Fay-Herriot Model for share below middle-class line, and average consumption in 2011 PPP Below Middle Average Class Cons. (1) (2) Population count 1.0000** Air temperature max 0.9071*** (0.0000) (0.0270) Grid Sum GDP 1.0000*** Max precipitation 1.0278*** (0.0000) (0.0104) Grid GDP Min 1.2266*** Nighttime lights 1.0000*** (0.0940) (0.0000) Vegetation Index max 1.0000** Grid GDP Min 0.0142*** (0.0000) (0.0225) Max precipitation 0.9987*** No local service road max value 1.0000*** (0.0004) (0.0000) Vegetation Index average 0.9999*** Grid GDP Mean 1.5187*** (0.0000) (0.1485) No local service road max 1.0000*** Population count 1.0000** (0.0000) (0.0000) Grid GDP Sum 1.0002*** (0.0001) 500m Elevation 0.9977*** (0.0006) Constant 2.8220*** Constant 9,759*** (0.1133) (5755) Observations 331 Observations 331 R-squared 0.4902 R-squared 0.5593 *** p<0.01, ** p<0.05, * p<0.1; selected *** p<0.01, ** p<0.05, * p<0.1; selected country/region dummy variables not shown country/region dummy variables not shown 75    Annex I – Sensitivity to Model Selection Technique Sensitivity analysis was conducted with respect to the method of automated model selection used in the FH application. The main alternative assessed was the lasso technique implemented using the lassopack suite of commands in Stata (Ahrens et al., 2018). In contrast to stepwise approaches which iteratively add variables to the model and assess the p-value of the indicator (keeping the variable in the model if it is below a certain threshold), the lasso approach minimizes the residual sum of squares subject to a constraint on the absolute size of coefficient estimates. There are thought to be two major advantages of lasso approaches over least squares. First the lasso tends to produce sparse solutions and thus facilitates model interpretation. Secondly, lasso can at times outperform least squares in terms of prediction due to lower variance. However, the stepwise models adopted in this paper were already quite sparse, and there were no great advantages observed in terms of minimizing the variance. Each of three lasso options are contrasted with simple step-wise selection. In no case are the differences remarkable. In this context, stepwise was maintained as the preferred approach due to its greater replication simplicity and ease of explanation to non-expert audiences. Figure 42: Lasso AIC vs. Stepwise 1 R² = 0.9445 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 43: Lasso EBIC vs. Stepwise 1 R² = 0.9232 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 Figure 44: Lasso BIC vs. Stepwise 1 R² = 0.9232 0.9 0.8 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 77    Annex J – K-fold Model Selection Sensitivity Analysis An additional concern of automated variable selection techniques is that overfitting will generate spurious models that perform poorly in out-of-sample prediction. One method of assessing the risk of overfitting is to train models while withholding a portion of the data and checking that the automated model development technique performs similarly well despite the missing data. This assessment is described in this annex (J), and the results (presented graphically in figures 45-49) show that there is little change in the estimates of interest when withholding a subset of data to validate performance and the presence of overfitting. In practice, this assessment was conducted by splitting the sample into 5 randomly assigned groups (called “folds”), and using 4 folds to develop the model, imputing by using the FH model to adjust the missing observations. Figure 45: Fold A Full Sample (Left); Withheld Sample (Right) 1 1 R² = 0.983 R² = 0.9795 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 78    Figure 46: Fold B Full Sample (Left); Withheld Sample (Right) 1 1 R² = 0.9784 R² = 0.9782 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Figure 47: Fold C Full Sample (Left); Withheld Sample (Right) 1 1 R² = 0.9754 R² = 0.9893 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 79    Figure 48: Fold D Full Sample (Left); Withheld Sample (Right) 1 1 R² = 0.9809 0.9 0.9 R² = 0.9804 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 Figure 49: Fold E Full Sample (Left); Withheld Sample (Right) 1 1 R² = 0.978 R² = 0.9693 0.9 0.9 0.8 0.8 0.7 0.7 0.6 0.6 0.5 0.5 0.4 0.4 0.3 0.3 0.2 0.2 0.1 0.1 0 0 0 0.2 0.4 0.6 0.8 1 0 0.2 0.4 0.6 0.8 1 80    To access full collection, visit the World Bank Documents & Report in the Poverty & Equity Global Practice Working Paper series list. www.worldbank.org/poverty