78168 Report No: Kyrgyz Republic: Minimum Living Standards and Alternative Targeting Methods for Social Transfers A Policy Note1 June, 2013 Poverty Reduction and Economic Management Unit Europe and Central Asia Region Document of the World Bank 1 The findings, interpretations, and conclusions expressed in this paper are of authors only. They do not necessarily represent the views of the the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. 1 Kyrgyz Republic: Minimum Living Standards and Alternative Targeting Methods for Social Transfers Franziska Gassmann June 2013 Table of Content Executive Summary ............................................................................................................................. 3 Introduction......................................................................................................................................... 5 Minimum living standards ................................................................................................................... 5 National poverty lines ..................................................................................................................... 6 Minimum Subsistence Level ............................................................................................................ 9 Guaranteed Minimum Income ...................................................................................................... 10 Minimum wage and basic pension ................................................................................................ 10 Why are the standards so different? Comparing the MSL with the FPL ....................................... 11 How can we explain the very low extreme poverty rate and the number of MBPF beneficiaries? ....................................................................................................................................................... 12 Could the GMI be linked to the MSL? ........................................................................................... 14 Modeling alternative targeting methods, thresholds and benefits .................................................. 15 Categorical targeting ..................................................................................................................... 16 Targeting by proxy-means test ...................................................................................................... 21 Current methodology with increased GMI.................................................................................... 26 Cost efficiency of alternative targeting options ............................................................................ 30 Conclusion ......................................................................................................................................... 31 References ......................................................................................................................................... 33 Annex 1: Model for proxy-means test............................................................................................... 34 Annex 2: Alternative models for modeling benefit take-up.............................................................. 38 Annex 3: Cost efficiency extreme poverty ........................................................................................ 40 2 Executive Summary The objective of this policy note is to analyze and discuss the links between different minimum living standards currently used in the Kyrgyz Republic, and to analyze the potential of improving the targeting performance and adequacy of the Monthly Benefit for Poor Families (MBPF) by changing the targeting method and/or increasing the benefit levels. The Kyrgyz Republic uses various minimum living standards, each serving a specific purpose: national poverty lines, Minimum Subsistence Level, Guaranteed Minimum Income, minimum wage and basic pension level. National poverty lines (absolute and extreme poverty line), which are annually calculated by the National Statistical Committee (NSC) based on the Kyrgyz Integrated Household Survey (KIHS), are used to assess the nature and extent of poverty in the country and the distributional outcomes of social protection policies. Poverty is an outcome and identifies those households with average consumption below the poverty line. Poverty analysis is essential for the design of appropriate public policies aimed at supporting the poorest and most vulnerable households. It is vital for the evaluation of the effectiveness and efficiency of public policies. The two national poverty lines are calculated based on the actual consumption of the population. The extreme poverty line identifies those households and individuals with consumption levels below the minimum required to satisfy a daily calorie intake of 2100 kcal per person. The absolute poverty line includes an allowance for non- food goods and services deemed necessary to satisfy other basic needs. The Minimum Subsistence Level (MSL) is a normative standard governed by Kyrgyz Law. Its food component reflects a balanced diet providing 2100 kcal per person per day. The main purpose of the MSL is to assess the overall living standard of the population, to identify social policy interventions and to determine minimum state labor guarantees. The composition of the MSL is reviewed every five years, while its value is adjusted quarterly by the NSC based on registered prices. The Guaranteed Minimum Income (GMI) is a budget-driven threshold to determine eligibility for the MBPF, a means-tested cash transfer targeted at poor families with children. Families are eligible for the transfer if the family income is below the GMI. The level of the GMI is adjusted on an irregular basis as it depends on available government resources. As the levels of the various minimum standards differ considerably (see table below), the share of the population with average consumption below the thresholds varies as well. Most notably, based on average consumption per capita (the main indicator to assess the poverty status of a household), no individual has a living standard below the GMI. This seemingly contradicts the finding that 11 percent of the population benefits from the MBPF. The main reason lies in the difference between administrative family income as used for MBPF eligibility and actual household consumption, which is used to assess the policy outcome. Absolute PL Extreme PL MSL (total) MSL (food) GMI Monthly value (KGS) 1,745 1,050 3,502 2,276 310 Poverty rate (%) 33.7 5.3 86.9 58.3 0.0 Source: KIHS 2010 Furthermore, the level of the GMI is low compared to the extreme poverty line. While it covered 50 percent of the extreme poverty line at its introduction in 1998, the GMI reflected only 28 percent of 3 the extreme poverty line in 2011. The Law governing the GMI does not foresee in any regular indexation mechanism, contrary to the MSL or the basic pension. Linking the GMI to the MSL would offer a solution. From an economic and social perspective the MSL can be considered as a future welfare benchmark. Defining the GMI as a percentage of the MSL would guarantee its regular adjustment over time. Setting the GMI initially at 20-30 percent of the food component of the MSL for children of the previous year is financially viable. Over time this share can be increased closing the gap with the extreme poverty line. The question remains, whether an increase of the GMI would improve the targeting performance and poverty-reduction capacity of the MBPF. In principle, a higher GMI increases the number of eligible families with children, and potentially, the number of actual beneficiaries. As long as the GMI is below the extreme poverty line, a reduction of the poverty rate is not given by definition. However, one could expect a further reduction of the poverty gap, which means that the extremely poor become less poor. Abandoning the concept of the GMI Change in extreme poverty gap altogether and moving to a different Extreme poverty gap (%) 1.0 0.8 targeting methodology may lead to better 0.6 results given the current budget (option 0.4 low 0.2 med ‘low’ in the figures). For that purpose, the 0.0 high potential of two alternative targeting methods are evaluated against a higher GMI. The first method considers a categorical approach, whereby families Change in extreme poverty rate with several children are eligible for the 6 Extreme poverty rate (%) transfer irrespective of the level of the 4 low family income. Analysis of the KIHS shows 2 med that children have an above proportionate 0 high risk of living in poverty compared to the average population. Furthermore, families with several children are especially disadvantaged. For the second alternative, the current means-test (assessment of total family income) is replaced by an approximate estimation of the living standard of the household. This is a so-called proxy means-test. Based on the KIHS, the relationship between low consumption levels and easy-to-observe indicators is estimated. Indicators, such as the demographic composition of the household, housing living conditions, possession of assets, employment and education status of household members, or location are used to identify poor households. The empirical analysis gives a score for each indicator. Counting the total score for an applicant household will determine whether a household is eligible for a transfer or not. Comparing the three alternatives (higher GMI, benefit for families with several children, proxy means-test) shows that the proxy means-test would achieve better outcomes in terms of targeting performance and extreme poverty reduction for the same amount of money. It would reduce both the exclusion and inclusion error. However, changing the current targeting methodology may entail considerable costs, especially during start-up. Furthermore, the final outcomes may be different from the simulated results as not all households may apply for a transfer. 4 Introduction One of the objectives of the Strategy for Social Protection Development 2012-2014 of the Government of the Kyrgyz Republic (GoKG) is to improve the targeting performance of the Monthly Benefit for Poor Families (MBPF). The MBPF is the only last-resort social transfer program in the Kyrgyz Republic aimed at providing a guaranteed minimum income to the population. Currently, eligibility for the MBPF is determined by a means-test. Moreover, only families with children can apply. A second objective of the strategy, which is closely linked to the first, aims at improving the poverty reduction effectiveness of the transfer. The size of the MBPF is defined by the difference between family income per capita and the level of the Guaranteed Minimum Income (GMI). Currently, the level of the GMI is determined by the funds allocated by the Ministry of Finance, which depend on the available government resources. The share of funds allocated to social transfers is not fixed. The MSD would like to revisit the GMI and investigate whether it could be tied to another standard used for policy making in the Kyrgyz Republic. This link should ensure the annual adjustment of the GMI, protect its real value and reduce the burden of frequent ad-hoc recalculations of benefits for the social workers at the local level. Furthermore, the MSD considers piloting alternative targeting methods that could further improve the targeting performance of the MBPF. The objective of this report is twofold: first, analyze and discuss the linkages between different minimum standards currently used in the Kyrgyz Republic and with the MBPF, and, secondly, to analyze the potential of alternative targeting methods and benefit levels. The analysis in this report mainly draws on data from the Kyrgyz Integrated Household Survey (KIHS) 2010. The report is structured as follows: the next section analyzes and compares different minimum living standards as currently used in the Kyrgyz Republic. Subsequently, we simulate different targeting options using static micro-simulation. Categorical targeting, proxy-means testing and the current means-test given a higher GMI are analyzed and eventually compared in terms of cost efficiency. The last section concludes. Minimum living standards Minimum living standards can be defined in various ways depending on the purpose they serve. Poverty lines, for example, are used to measure the outcomes of poverty reduction policies. They are derived empirically and measure living standards in terms of household consumption. Other standards are based on normative decisions, such as most minimum subsistence levels used throughout Eastern Europe and Central Asia. Other standards define entitlements such as minimum wages and pensions. This section tries to disentangle some of the questions frequently raised by the GoKG. First, the different standards are shortly described. Subsequently, we compare and discuss different related standards with each other raising questions, such as, why the Minimum Living Standard and the Food Poverty Line are so different even though both assume a minimum daily calorie intake of 2,100 kcal, and whether we can explain the very low extreme poverty rates and the actual share of MBPF beneficiaries. 5 National poverty lines The main purpose of a national poverty line is to analyze they extent and nature of poverty in the country, and to assess the (distributional) outcomes of public policies. More specifically, national poverty lines are used to analyze the targeting performance and poverty reduction impact of social protection policies, such as the Monthly Benefit for Poor Families (MBPF) and other non-contributory social transfers (MSB, categorical state benefits, energy compensations). The national poverty lines used in the Kyrgyz Republic are derived from nationally representative household survey data. The Kyrgyz Integrated Household Survey (KIHS) collects annually information from more than 5,000 households, which are selected by a 2-stage random sampling process. The KIHS collects detailed information on income and consumption, the demographic composition of the household, labor market participation, assets, housing, land and livestock possession. The data on household expenditure and consumption is then used to analyze consumption patterns of households and to establish two poverty lines, which are set at values covering a minimum consumption basket. Food poverty line The objective of the food poverty line is to identify the extremely poor households and individuals in the society. Basically, it is set at the monetary value covering the costs of obtaining 2,100 kcal per person per day.2 Contrary to normative decisions on what a minimum diet of 2,100 kcal should contain, the food poverty line is empirically derived from the KIHS data. It reflects the actual food consumption habits of the Kyrgyz population. More specifically, it looks at food consumption of low- income households belonging to the second to fifth consumption deciles, excluding the poorest ten percent and the richest 40 percent of households.3 Based on the analysis of the KIHS 2011, Tsirunyan (2012) finds that 86 different food items account for 97 percent of total food consumption. Using their respective caloric values and shares in total food consumption, the costs of the minimum food basket is established. Table 1 presents the composition of the minimum food consumption basket in terms of caloric value and costs to obtain a daily minimum food intake of 2,100 kcal per person. Absolute poverty line In addition to food, households also have other basic needs, such as shelter, clothing, health and education. The absolute poverty line reflects these additional needs and includes an allowance for non-food goods and services. In order to estimate the non-food allowance, the reference group includes those households with food consumption near the food poverty line (10 percent above and below). The share of non-food consumption in total consumption of the reference group determines the non-food allowance to be added to the food poverty line. Based on the analysis of the KIHS 2011, total household consumption of the reference group consists of 62 percent food and 38 percent non- food goods and services (Tsirunyan, 2012). 2 The standard of 2,100 kcal, determined by the GoKG, is in line with international standards developed by WHO or FAO. 3 Previously, the reference group covered households from deciles 3-5 (GoKG, Resolution no. 115). For more information, see also Tsirunyan (2005, 2012). 6 Table 1. Composition of the minimum food basket by caloric value and costs, 2011 Share of 2,100 kcal (%) Share of costs (%) Bread and cereals 58.1 27.1 Milk and dairy products 4.8 8.6 Meat and meat products 4.4 19.3 Fish and fish products 0.0 0.2 Vegetable oil, margarine and other fats 13.7 7.1 Eggs 0.6 1.6 Potatoes 3.9 4.7 Vegetables, melons and gourds 3.1 11.1 Fruits and berries 1.7 5.2 Sugar 8.9 12.1 Tea, coffee, cacao 0.2 1.5 Non-alcoholic beverages 0.2 0.4 Other food products 0.3 1.1 Total 100 100 Source: Tsirunyan (2012, p.12) Figure 1. Development of the value of the food and absolute poverty line, 1998-2011 2500 2000 KGS per month 1500 1000 500 0 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Food poverty line Absolute poverty line Development of the national poverty lines over time Although the KIHS is collected every year, the poverty lines are not re-calculated annually. This would make the analysis of poverty developments over time difficult as the standards would change every year. Normally, once the poverty lines have been calculated, their values are adjusted annually using the consumer price index. Figure 1 shows the development of the value of the national poverty lines since 1998. Since 1996, poverty lines have been recalculated four times (see GoKG, 2011 and Tsirunyan 2012): - In 2000 due to the expansion of the sample size from 2,000 to 3,000 households; - In 2003 due to the integration of the Household Budget and Labor Force Surveys; - In 2008 and 2011 due to the average annual inflation having exceeded ten percent. 7 The vertical lines in figure 1 indicate the years that the poverty lines were recalculated. The last two revisions had a larger impact than previous ones. Figures 2 and 3 present the composition of the minimum food basket in terms of caloric values and costs over the years. The main caloric value is still derived from bread and cereal products, followed by fats and sugar. Changes over the three years have been minimal. The changes are slightly more pronounced with respect to the cost composition of the minimum food basket. Over time, the share of the minimum food basket in the absolute poverty line has slightly increased from 60 percent in 1998 to 62 percent in 2011 (table 2). Figure 2. Composition of minimum food basket by caloric value, 2003-2011 (Source: GoKG 2011, Tsirunyan 2012) 100% 7 10 8.9 90% Other food products 5 04 3.9 Non-alcoholic beverages 80% 11 12 13.7 Tea, coffee, cacao 70% 3 5 04 4.4 Sugar 60% 04 4.8 Fruits and berries Vegetables, melons and gourds 50% Potatoes 40% Eggs 30% 64 Vegetable oil, margarine and other fats 59 58.1 Fish and fish products 20% Meat and meat products 10% Milk and dairy products Bread and cereals 0% 2003 2008 2011 Figure 3. Composition of minimum food basket by costs, 2003-2011 (Source: GoKG 2011, Tsirunyan 2012) 100% 90% 9 10.5 12.1 Other food products 80% 11 Non-alcoholic beverages 9.9 11.1 Tea, coffee, cacao 70% 6 4.5 4.7 Sugar 60% 9 7.7 7.1 Fruits and berries Vegetables, melons and gourds 50% 12 16.6 19.3 Potatoes 40% 7 Eggs 8.7 8.6 Vegetable oil, margarine and other fats 30% Fish and fish products 20% 37 Meat and meat products 31.9 27.1 Milk and dairy products 10% Bread and cereals 0% 2003 2008 2011 8 Table 2. Composition of the absolute poverty line, 1998-2011 1998 2003 2008 2011 Share of minimum food basket 60.2 62.9 63.9 62.2 Share of non-food goods and services 39.8 37.1 36.1 37.8 Source: GoKG 2011, Tsirunyan 2012 Table 3. Composition of the MSL per Government Resolution No. 694, 2009 (%) Average Working age Average Per age group per capita adult Pensioner child 0-7 7-14 14-17 Food 65 61 70 71 71 72 71 Non-food goods 16 17 10 16 15 16 18 Services 17 19 20 13 14 12 11 Taxes 2 3 0 0 0 0 0 Total 100 100 100 100 100 100 100 Monthly value in 2011, KGS 4390.02 4920.71 3932.23 3708.79 3278.86 3867.65 4198.87 Table 4. Composition of MSL in kilograms and value, average per month per capita Based on 2009 resolution For 2011 in kg Share of food Value (KGS) Share of food Total food, of which 106.41 100 2853.52 100 Bread and cereals 18.83 17.7 416.24 14.6 Milk and dairy products 45.33 42.6 525.56 18.4 Meat and meat products 6.32 5.9 856.32 30.0 Fish and fish products 0.35 0.3 56.5 2.0 Vegetable oil, margarine and other fats 1.1 1.0 85.54 3.0 Eggs (16 * 60g) 0.96 0.9 105.63 3.7 Potatoes 4.69 4.4 100.73 3.5 Vegetables, melons and gourds 13.66 12.8 239.92 8.4 Fruits and berries 13.27 12.5 332.93 11.7 Sugar 1.66 1.6 103.2 3.6 Tea, coffee, cacao 0.07 0.1 28.42 1.0 Other food products 0.17 0.2 2.53 0.1 Non-food goods 702.4 Services 746.3 Taxes 87.8 Total 4,390.02 Source: Ministry of Social Development Minimum Subsistence Level The Minimum Subsistence Level is governed by the Law on Guaranteed Minimum Social Standards of May 26, 2009, No. 170.4 It is based on a normative consumption basket, whose composition is determined by a Working Group consisting of the Ministries of Health, Social Development, 4 Prior to 2008, it was called Minimum Consumer Basket. 9 Economic Regulation and the National Statistical Committee. The Ministry of Health leads the discussion on the food component with respect to minimum caloric requirements and nutritional values. The Ministry of Economic Regulation is responsible for forecasting the MSL. Prior to 2006, the minimum consumer basket has been reviewed only every 10 years (1995, 2006). The new Law requires a review every 5 years instead. The latest revision took place in 2009. In between, the MSL is quarterly updated by the National Statistical Committee using consumer price indices based on registered prices. The MSL is established for different population groups (working age men, working age women, pensioners, and children of different age) and per oblast (table 3). The MSL is used to assess the living standards of the population, to identify social policy interventions and to determine minimum state labor guarantees. The current MSL contains 32 different food items, accounting for 2,101 kcal per average person per day.5 Food accounts for 65 percent of the total MSL in 2011. Milk and dairy products account for the largest share in terms of quantity (kg) of food, followed by bread and cereals, vegetables and fruits. In terms of monetary value, meat and meat products account for 30 percent of the total food basket (table 4). Guaranteed Minimum Income The Guaranteed Minimum Income (GMI) is a budget-driven threshold used to determine eligibility for the Monthly Benefit for Poor Families (MBPF), which is a means-tested cash transfer targeted to poor families with children. It is governed by the Law on State Benefits from December 29, 2009, No. 318. Prior to 2010, it was called Guaranteed Minimum Consumption Level (GMCL) and was in addition used to determine the level of the Monthly Social Benefit (MSB), a categorical state transfer for different beneficiary groups, such as individuals disabled since birth and pensioners without pension entitlements. At its introduction in 1998, the value of the GMI (GMCL) was set at 50 percent of the extreme poverty line. However, there was no provision in the law officially tying the GMI to the extreme poverty line. The level of the GMI depends on the funds allocated by the Ministry of Finance and as such is entirely dependent on the available government resources. The value of the GMI has eroded over time. Notwithstanding several ad-hoc adjustments over the years, the value of the GMI never reached the initial value of 50 percent of the extreme poverty line again as at the time of its introduction. In 2011, the value of the GMI was 8 percent of the MSL and 28 percent of the extreme poverty line (figure 4). Minimum wage and basic pension The official minimum wage is governed by the Law on Minimum Wages in the Kyrgyz Republic from October, 12, 2008, No. 210 and the Law on the Republican Budget for 2011 and Forecast for 2012- 2013 from March 30, 2011, No. 8. It is used to determine the guaranteed monthly salary for full-time unqualified labor under normal working conditions. The minimum wage is not used for the calculation of social transfers. However, if a worker entering retirement has no salary records, the minimum wage is used to calculate his/her pension entitlements. The minimum wage was set at KGS 690 per July 1, 2011, which represents 16 percent of the average MSL. The basic pension is set in relation to the average salaries in the country. The Law on State Pension Social Insurance of July 21, 1997, No. 57 and the President’s Decree on the Increase of the Basic Pension determine that the basic pension cannot be less than 12 percent of the average salaries in 5 In 2006, the food basket provided for 2,431 kcal per average person per day. 10 the previous year. In order to be eligible for a full basic pension, women need to have contributed during 20 years, and men during 25 years. Per 1 October 2011, the full basic pension has been increased with 500 KGS to 1,500 KGS per month (21 percent of average salaries in 2010). Figure 4. Development of the GMI (GMCL), 1998-2011 1600 1400 1200 1000 KGS per month 800 600 400 200 0 1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011 Guaranteed minimum consumption level (GMCL) som a month Food poverty line Why are the standards so different? Comparing the MSL with the FPL Although the food baskets of both the food poverty line (FPL) and the MSL are based on a daily minimum of 2,100 kcal per person, the composition of the baskets and their costs differ significantly. According to the MSL, a monthly food basket costs 2,853 KGS on average per person, while the value of the food basket empirically derived for the FPL is 1,340 KGS per person for 2011. This is less than 50 percent of the normative MSL value. Figure 5 contrasts the composition of the two minimum food standards. Note, that the first two stacked bars in figure 5 are not directly comparable. The composition of the FPL is based on the caloric share of the different products, while the MSL is expressed in kilograms. The actual food consumption habits of the FPL reference population (deciles 2-5) deviates significantly from the normative MSL basket. Bread and cereals deliver almost 60 percent of the daily calories, followed by fat and sugar. Milk and dairy products, which have a prominent place in the MSL, account for only five percent in actual consumption. It is important to keep in mind that the food basket of the MSL is based on normative standards representing a well- balanced and healthy diet, both in terms of caloric and nutrient value (vitamins, minerals, proteins, etc.). The difference in the total costs of the two baskets can be explained by the difference in actual versus normative consumption and the prices used to calculate the costs. The FPL basket is entirely based on price information as provided by households in the KIHS, while the MSL is based on registered prices. The fact that most households get some part of their food from own production or by hunting, fishing and gathering may have an impact on the reported prices of food items in the KIHS. Their price information may be biased. 11 Figure 5. Composition of the food poverty line and MSL, 2011 100% 02 8.9 04 90% 1.7 12 12.1 12 3.1 Other food products 80% 5.2 13 08 Non-alcoholic beverages 13.7 11.1 70% Tea, coffee, cacao 4.4 01 03 06 Sugar 60% 4.8 7.1 Fruits and berries 50% 30 Vegetables, melons and gourds 19.3 40% 43 Potatoes 30% 58.1 8.6 Eggs 18 Vegetable oil, margarine and other fats 20% 27.1 Fish and fish products 10% 18 15 Meat and meat products 0% Milk and dairy products Food PL (kcal) MSL (kg) Food PL MSL Bread and cereals kcal or kg share share total costs Comparing the two stacked bars on the right in figure 5, we notice that the differences are less pronounced when considering the share of costs of different items in the total budget. Based on the MSL the largest post in the total costs is taken up by meat and meat products, followed by milk and dairy products and bread and cereals. Due to the large share of bread and cereals in the FPL basket, these items also account for the largest part in the total budget required to obtain the FPL basket. Meat, sugar and vegetables are the next largest items in the FPL budget. Comparing the composition of the calories (kilograms) with the composition of the costs provides an indication about the price levels of the different food items. While consumption of bread and cereals account for nearly 60 percent of the total caloric value of the FPL basket, it accounts for only one quarter of its total cost. On the other hand, meat and meat products deliver less than five percent to the minimum calorie intake, they account for 20 percent of the costs of the minimum basket. Summarizing, the large difference between the costs of the two minimum food baskets stems from the difference between what households actually consume and the normative standards underlying the MSL. The difference between observed and registered prices used to calculate the costs of the two baskets may explain part of the difference, too. How can we explain the very low extreme poverty rate and the number of MBPF beneficiaries? The proportion of the population living with less than the various minimum standards differs greatly across the different measures (table 5). 87 percent of the Kyrgyz population is living with less than the total MSL (average value per person) and 58 percent does not meet the MSL food standard. Based on the empirically derived poverty lines, 34 percent is living in absolute poverty and just five percent of the population consumes less than the food poverty line. If we take the GMI as threshold, no one can be identified as having less. However, based on benefit incidence analysis, 11 percent of the population is living in a household receiving the MBPF. How can this paradox been explained? 12 Table 5. Poverty rates based on different standards, 2010 Absolute PL Food PL MSL (total) MSL (food) GMI Monthly value (KGS) 1,745 1,050 3,502 2,276 310 Poverty rate (%) 33.7 5.3 86.9 58.3 0.0 Source: KIHS 2010 Poverty analysis is primarily used to measure the scope and nature of poverty and to analyze the outcomes of various public policies. It identifies those groups of the population that are living below the absolute or food poverty line. The living standard of a household is determined by its total household consumption, excluding expenditures for taxes, alimonies, agricultural production and transfers to other households. Total household consumption includes expenditures for food, non- food goods and services and accounts for the possession of durable goods. It also includes consumption of goods from own production, especially food. Household consumption is the preferred welfare indicator as it better reflects the actual living standard of a household. It is less sensitive to (seasonal) variations in income because households smooth consumption over time taking into account expected income.6 Furthermore, households are usually more truthful with respect to the information provided in household budget surveys, especially if they are required to keep a diary on consumption and expenditures as part of the survey. Information on income is less reliable in general, as people have a tendency to underreport income. Total household consumption is further divided by the number of household members without taking into account economies of scale or the demographic composition of the household.7 The average household consumption per capita is then compared to the poverty line. The guaranteed minimum income level (GMI) has an entirely different purpose. As described above, it is set by the GoKG based on the available budgetary resources. Its main purpose is to determine whether a household is eligible for the Monthly Benefit for Poor Families (MBPF) and define the level of the benefit. The eligibility assessment (means test) is based on household income. It includes all types of income, both formal and informal. Households have to submit proof for income from formal sources. In order to account for land ownership, income from land is imputed using regional coefficients. While it is rather straightforward to control formal income, informal income, such as from informal work or remittances, or income from subsistence agriculture is more difficult to verify by the social worker. As already discussed above, the food poverty line represents the costs of obtaining a minimum food basket satisfying 2,100 kcal per person per day. It represents the costs for a household that has to purchase all food items either in shops or on the market. It does not take into account consumption from own production or from gathering or hunting. On average, food from own production accounts for 33.5 percent of total food consumption (figure 6, horizontal line represents the national average). The shares are higher in rural areas and lower for the richest households. 6 This is also called the ‘permanent income hypothesis’. 7 No equivalence scales are used. All measures are per capita. 13 Figure 6. Share of food from own production in total food consumption, 2010 (Source: KIHS 2010) 40 30 20 10 0 1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10 Consumption deciles Urban Rural Based on the KIHS, 3.9 percent of the population is living in a household with total household income below the GMI. The definition of ‘total household income’ derived from the survey is, however, not the same as ‘administrative income’ used to assess benefit eligibility. We therefore need to find a proxy for the administrative definition of income used for the eligibility assessments in the survey data. Table 6 estimates the proportion of the population having less than the GMI using different definitions of income. Excluding food in kind from total consumption has a negligible impact on the percentage below the GMI. Using total household income per capita results in four percent of the population below the eligibility threshold. We continue by deducting specific formal and informal transfers. The MBPF is deducted by definition as we need to understand eligibility prior to the receipt of the benefit. The MSB is deducted as it is excluded from the income test. We then subtract income from private transfers as these are difficult to verify. Six percent of the population is then identified as having less than the GMI. This is still less than the eleven percent that is living in MBPF recipient households. Table 6. Population below the GMI according to different income criteria, 2010 Income indicator (all per capita) Population below GMI (%) Total household consumption 0.0 Total household consumption, w/o food in kind 0.2 Total household income 3.9 Total household income, w/o MSB, MBPF 4.6 Total household income, w/o MSB, MBPF, private transfers 5.9 Source: KIHS 2010 Could the GMI be linked to the MSL? One of the issues related to the GMI is the erosion of its real value over time. Regular indexation is not foreseen under the current law. The MSL, on the other hand, is mainly used to forecast the 14 development of living standards in the country but does not play a role in targeting or setting of benefit levels. Yet, it is recalculated quarterly and as such follows the price developments in the country. From an economic and social development perspective the MSL could be considered as a welfare benchmark to be reached in the future. The guaranteed minimum income could initially be defined as a percentage of the MSL. Over time and given a growing economy, the GoKG could raise this percentage and slowly getting closer to the MSL. Linking the GMI to the MSL has its merits. It is a rather common practice in the region. Russia, Kazakhstan, Slovakia, Lithuania, Turkmenistan, and the Czech Republic are some of the countries using the MSL as the standard to determine social assistance eligibility thresholds. As the MSL is regularly recalculated, the GMI would be automatically adjusted. This could be done annually in order to reduce the burden of recalculating benefit entitlements. In the Czech Republic, for example, eligibility levels are only adjusted if prices have increased with more than 5 percent. The issue is at what level of the MSL the GMI should be set initially. As discussed above, the MSL includes both food and non-food items. For the purpose of tying the GMI to the MSL, the costs for the basic food package of the previous year could serve as reference. Furthermore, since the MBPF is targeted at children, the reference level is the costs of the food basket for children, which was 2,633 KGS per month per child in 2011. On average, 38 percent of food is from own production for the bottom 40 percent of the population (42 percent in rural areas for the bottom 40 percent). Setting the standard at 60 percent of the food-MSL for children would provide a GMI of 1,580 KGS in 2012 (table 7). This is more than four times the value of the GMI, which was raised to 370 KGS in 2011 (580 KGS in November 2012). From a budgetary perspective this level does not seem to be feasible. However, setting the GMI initially at 30 percent of the food basket of the MSL for children of the previous year would be financially more sustainable. With economic growth, the level of the GMI as a share of the MSL could be increased over time. But, this is an entirely arbitrary decision not based on any normative or empirical standard, but purely budget- driven. The MSD uses various options for their MBPF budget forecasts. So far, their proposals continue to link the GMI to the extreme poverty line. The ambition is to increase the level of the GMI to 100 percent of the food poverty line. Table 7. GMI values for different levels of MSL As a percentage of the food basket of the MSL for children 20% 30% 40% 50% 60% Potential GMI values (KGS) 527 790 1,054 1,362 1,580 Modeling alternative targeting methods, thresholds and benefits The current Monthly Benefit for Poor Families (MBPF) is the only social transfer specifically targeted at the poorest households. Its eligibility is based on average family income and the presence of children in the household. Average family income has to be below the GMI. The size of the transfer is the difference between the family income per capita and the GMI times the number of children up to 15 the age of 16 (or 21 if fulltime students). In 2010, 362 thousand beneficiaries have received the MBPF. The average transfer was 235 KGS per beneficiary per month.8 Table 8. MBPF beneficiaries and expenditures, administrative data, 2005-2011 2005 2006 2007 2008 2009 2010 2011est MBPF recipients (thds) 482 481 475 434 371 362 398 Average MBPF per month (KGS) 89 124 125 120 172 235 290 Total annual expenditures MBPF (mio KGS) 508 773 695 673 755 1330 1411 Source: Ministry of Social Development The outcome analysis based on the KIHS 2010 shows that the MBPF is indeed a progressive transfer predominantly allocated to poor households (table 9). Almost two thirds of total transfers are received by the poorest quintile. Of more concern are the size of the exclusion error and benefit adequacy as measured by the share of benefits in total household consumption in recipient households. Less than one in three individuals of the poorest quintile lives in a household receiving the MBPF. More than 70 percent of the poorest group are not covered by the transfer. The limited size of the transfer reduces its importance in total household consumption even for the poorest households. On average, the MBPF accounts for less than ten percent of total household consumption for the bottom quintile. As a result, the impact on poverty reduction is minimal. Recognizing the sizable exclusion error and the limited impact of the current MBPF on poverty reduction, the Ministry of Social Development (MSD) is exploring alternative ways of targeting the MBPF. Table 9. Targeting performance MBPF, 2010 Consumption quintile* I II III IV V All Coverage 28.9 10.8 9.0 4.7 2.3 11.2 Distribution beneficiaries 51.8 19.4 16.1 8.5 4.2 100.0 Distribution benefits 61.4 19.7 10.4 5.6 2.9 100.0 Share in total consumption (only recipients) 8.4 4.9 2.5 2.6 1.8 5.9 Note: Quintiles are based on annual per capita consumption before transfers, assuming a marginal propensity of 33 percent. Source: KIHS 2010. The next section will simulate two alternative targeting options and compare it with the current methodology given a higher GMI. The analysis uses data from the KIHS 2010. All results are based on the situation in 2010. For each method, we simulate a low, medium and high benefit scenario based on the budgetary parameters provided by the MSD. The low scenario is based on the 2010 annual budget allocation of 1,330 mio KGS. The medium scenario foresees in raising the budget to 2,000 mio KGS and the highest scenario would aim for an allocation of 5,600 mio KGS per year. Categorical targeting One alternative to the current means test is the allocation of the MBPF to specific categories of the population of whom we know that they have a very high risk of being extremely poor. Eligibility is based on, for example, demographic characteristics. Common examples for categorical transfers are child allowances or social pensions. The Monthly Social Benefit (MSB) is an example of a categorical transfer as its allocation does not depend on household income but on specific characteristics of the 8 Information provided by the Ministry of Social Development, February 2012). 16 beneficiaries, such as disability status. The disadvantage of categorical targeting is its inaccuracy. There will always be leakage to non-poor households. The higher the correlation between the identified categories and poverty, the lower the targeting errors. On the positive side, categorical targeting reduces the administrative burden both for the applicant and the policy maker. It can also easily be combined with other targeting methods. The first task is to define categories that allow distinguishing between the extremely poor and other households in order to limit errors of inclusion but also to ensure a high coverage of the extremely poor. Categories can be either based on demographic indicators, geographical location, housing characteristics, the possession of certain durable assets or any combination thereof. For geographical targeting poverty rates ideally need to be calculated at least at district level. This is not possible with the KIHS as the sample is only representative at oblast level. The search for other categorical indicators to identify the extremely poor at oblast level rendered demographic indicators as potential candidates for categorical targeting. As in most low and medium income countries, poverty in the Kyrgyz Republic is highly correlated with the presence of children in the household. The more children in a household, the higher the poverty risk for the household. The extreme poverty rate of children below the age of six is 9.4 percent compared to the national average of 5.3 percent. For the subsequent analysis we focus on the poorest 10 percent of the population. Households belonging to the poorest ten percent have on average 2.7 children below the age of 16, compared to 1.2 children in other households. The highest average number of children are observed among the poorest households in Osh and Talas oblast with 3.6 and 3.4 children on average per household (figure 7). Figure 7. Average number of children per household (Source: KIHS 2010) 04 Number of children per household 03 02 01 00 Decile 1 Decile 2-10 Individuals living in households with three or more children below 16 have a considerably higher risk of belonging to the poorest ten percent of the population (table 10). This risk increases with every additional child. Having very young children further exacerbates the extreme poverty risk. 17 Table 10. Profile of the poorest 10 percent of the population, percentages, 2010 Belonging to Share in total Belonging to Share in total poorest 10% population poorest 10% population Total 10.0 100 By age: Issykul 8.2 8.1 below 6 16.8 11.7 Jalal-Abad 15.2 18.9 6-15 11.0 22.0 Naryn 22.2 4.8 16-20 8.0 10.3 Batken 10.9 8.1 21-40 11.6 24.9 Osh 12.3 25.3 41-60 5.9 23.5 Talas 8.2 4.2 61-70 6.3 4.1 Chui 5.8 14.9 71 or older 7.7 3.5 Bishkek 1.4 15.7 Households with Households with No child <16 2.0 20.9 No child <6 4.4 54.4 1 child < 16 4.1 22.8 1 child <6 10.1 28.4 2 children < 16 9.2 27.5 2 children <6 21.4 13.9 3 children < 16 15.1 17.7 3 children <6 55.3 3.2 4 children < 16 18.8 8.0 4 or more children <6 0.0 0.1 5 children <16 57.6 2.6 6 or more children <16 91.4 0.5 Disabled child 13.8 1.1 Disabled adult 15.4 5.8 Note: The poorest 10% are identified based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). Source: KIHS 2010 Based on the analysis of extremely poor households, the number of children seems to be a good indicator for identifying the poorest households in the Kyrgyz Republic. Several options have been tested: e1 three or more children below age of 16 e2 three or more children below age of 16, of which at least one child below 6 e3 three or more children below age of 16, OR two or more children below age of 6 e4 same as e1, but for Naryn & Batken: 2 or more children e5 same as e2, but for Naryn & Batken: 2 or more children e6 same as e3, but for Naryn & Batken: 2 or more children<16, OR one or more below 6 The choice for households with three children or more below the age of 16 is guided by the poverty profile as presented in table 10. Variations include stricter and looser eligibility criteria. One option (e2) requires that at least one of the three children is below the age of 6, while option e3 would also cover households that currently have two children below the age of 6. The analysis also showed that the situation slightly differs in Naryn and Batken. In these two oblasts the poverty risk is already significantly higher for households with two or more children compared to other households in the oblast. Therefore, options e4, e5 and e6 allow for different eligibility criteria in Naryn and Batken. One of the main properties of categorical targeting is the fact that each household meeting the criteria is eligible for a transfer irrespective of its level of income. Tables 11 and 12 contain the coverage rates for each simulated option, given that indeed all eligible households would apply (i.e. assuming perfect implementation). Analyzing the predicted outcomes per consumption quintile, 18 coverage rates of the poorest quintile range between 50 and 72 percent, resulting in exclusion errors of 28 to 50 percent. The exclusion error is driven by those households that have fewer than three children (or two in Naryn and Batken, depending on the simulated option). The highest coverage would be achieved with option e6, where all households with three or more children below 16 (Naryn and Batken: two or more), or with at least two children below the age of 6 (Naryn and Batken: one child below 6) would qualify. The most generous option (e6) also leads to the highest inclusion error, where 63 percent of the beneficiaries would not belong to the poorest quintile. Option e2 – the strictest definition - would result in the lowest inclusion error (55 percent). Table 11. Simulated coverage and distribution rates of categorical targeting options, percentages Coverage Distribution of beneficiaries Quintile e1 e2 e3 e4 e5 e6 e1 e2 e3 e4 e5 e6 Q1 58.9 49.9 67.1 63.5 53.3 72.1 40.9 45.1 38.4 39.0 43.7 37.0 Q2 39.6 29.7 47.1 45.5 33.0 52.2 27.6 26.9 27.0 28.0 27.1 26.8 Q3 26.1 17.7 31.6 30.3 20.1 36.0 18.1 15.9 18.0 18.6 16.4 18.4 Q4 13.3 9.4 20.1 15.9 11.0 23.7 9.2 8.5 11.5 9.8 9.0 12.2 Q5 6.0 4.0 8.9 7.5 4.7 11.1 4.2 3.6 5.1 4.6 3.8 5.7 total 28.8 22.2 35.0 32.6 24.4 39.0 100 100 100 100 100 100 Excl. error 41.1 50.1 32.9 36.5 46.7 27.9 Incl. error 59.1 54.9 61.6 61.0 56.3 63.0 Note: Quintiles are based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). Exclusion error = 100% - coverage Q1; Inclusion error = 100% - beneficiaries belonging to Q1. Source: KIHS 2010 Since the objective of the MBPF is to target the extremely poor households, table 12 only considers the poorest ten percent of the population and provides coverage rates by oblast. Coverage is higher in poor oblast and lower in the wealthier oblast such as Bishkek and Chui. For Naryn and Batken the less stringent eligibility criteria result in significantly higher coverage rates for options e4 to e6 compared to e1 to e3. In some oblasts, such as Osh and Talas, coverage rates could reach almost 90 percent among the poorest decile. Table 12. Simulated coverage of the poorest ten percent, by oblast, percentages e1 e2 e3 e4 e5 e6 Issykul 65.6 61.0 70.6 65.6 61.0 70.6 Jalal-Abad 62.8 46.0 76.1 62.8 46.0 76.1 Naryn 40.5 37.5 49.6 79.2 70.1 87.3 Batken 38.6 30.8 53.7 65.0 51.9 72.4 Osh 86.1 79.9 88.4 86.1 79.9 88.4 Talas 85.9 75.7 94.3 85.9 75.7 94.3 Chui 31.6 23.6 47.0 31.6 23.6 47.0 Bishkek 6.9 6.9 12.9 6.9 6.9 12.9 Note: Decile is based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). It identifies the poorest ten percent at the country level. Source: KIHS 2010 For the further simulation, we select the options with the lowest inclusion error (e2) and the lowest exclusion error (e6). For each of these two categorical targeting options, we simulate six benefit scenarios, where flat benefits are either allocated for each child below the age of 16 or per household. The benefit levels are derived from the budget forecasts and the number of eligible children or households based on the KIHS 2010. 19 Table 13. Benefit scenarios for simulation of categorical transfers Low Medium High Budget: 1,330 Budget: 2,000 Budget: 5,600 mio KGS mio KGS mio KGS Eligible children Monthly transfer per child below 16 (KGS) (#) E2: households with three or more children below 673,387 165 248 693 age of 16, of which at least one child below 6 E6: households with three or more children below age of 16, or two or more children below age of 6; 1,083,137 102 154 431 for Naryn & Batken: 2 or more children<16, or one or more below 6 Eligible Monthly transfer per household (KGS) households (#) E2: households with three or more children below 190,829 581 873 2445 age of 16, of which at least one child below 6 E6: households with three or more children below age of 16, or two or more children below age of 6; 364,939 304 457 1279 for Naryn & Batken: 2 or more children<16, or one or more below 6 Note: number of eligible children and households derived from KIHS 2010. Budget information from MSD. Table 14 summarizes the main results of the poverty impact simulation of the six benefit scenarios. The current poverty rates (before and after MBPF) are reported as well. Under the low benefit scenario, the impact on absolute poverty remains very small. To some extent, differences are only noticeable at two decimals. The higher the transfers, the larger the effect on poverty rates. From the poverty reduction perspective, it hardly matters whether the transfer is designed as a flat transfer per child or per eligible household. However, in terms of benefit distribution and leakage, allocating flat benefits to each child in eligible households results in a slightly more progressive distribution (table 15). Table 14. Simulated impact on poverty, categorical options E2 and E6 Absolute poverty Extreme poverty Rate % Gap % Rate % Gap % Before MBPF 33.93 7.55 5.63 0.84 After MBPF 33.68 7.43 5.34 0.79 E2: benefit per child low 33.32 7.31 5.40 0.76 med 33.03 7.20 5.14 0.72 high 31.72 6.61 3.79 0.56 E2: benefit per hh low 33.26 7.33 5.40 0.77 med 33.16 7.21 5.39 0.74 high 31.73 6.65 4.01 0.61 E6: benefit per child low 33.27 7.34 5.47 0.77 med 33.12 7.24 5.28 0.74 high 31.89 6.69 4.04 0.58 E6: benefit per hh low 33.40 7.36 5.48 0.78 med 33.16 7.26 5.28 0.76 high 32.17 6.76 4.12 0.61 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 20 Table 15. Simulated distribution of benefits, categorical options E2 and E6, percentages e2 e6 Quintile Flat benefit per child Flat benefit per hh Flat benefit per child Flat benefit per hh Q1 44.4 41.4 37.5 33.4 Q2 27.4 28.3 26.4 26.4 Q3 16.1 17.2 19.5 20.6 Q4 8.2 8.7 10.4 11.9 Q5 4.0 4.4 6.2 7.6 Total 100 100 100 100 Note: Quintiles are based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). Source: KIHS 2010 Categorical transfers to children not only benefit children but all individuals in eligible households (table 16). Compared to the current situation, the impact on poverty would only increase under the medium and high scenario. Given the budget of 2012, categorical transfers do not outperform the current means-tested MBPF with respect to the extreme poverty rate. However, they seem to be more effective in reducing the poverty gap (lower panel in table 16). Table 16. Simulated impact on poverty for categorical options E2 and E6, by age E2 (flat benefit per child) E6 (flat benefit per child) Before After MBPF MBPF low medium high low medium high Extreme poverty rate (%) below 6 10.06 9.44 9.60 9.07 6.39 9.69 9.60 7.03 6-15 6.12 5.54 5.65 5.15 3.71 5.80 5.41 4.06 16-20 4.51 4.35 4.40 4.24 2.96 4.51 4.15 3.11 21-40 6.87 6.58 6.66 6.42 4.77 6.70 6.60 4.97 41-60 2.90 2.90 2.84 2.83 2.26 2.88 2.78 2.37 61-70 2.89 2.89 2.89 2.89 2.87 2.89 2.89 2.87 71 or older 3.81 3.81 3.81 3.56 2.60 3.81 3.43 2.46 Extreme poverty gap (%) below 6 1.59 1.46 1.39 1.30 0.92 1.45 1.37 1.03 6-15 0.92 0.84 0.81 0.75 0.56 0.83 0.79 0.60 16-20 0.45 0.44 0.41 0.39 0.34 0.41 0.39 0.30 21-40 1.07 1.00 0.96 0.91 0.72 0.98 0.94 0.75 41-60 0.44 0.43 0.41 0.39 0.33 0.41 0.40 0.33 61-70 0.30 0.30 0.30 0.30 0.30 0.29 0.29 0.26 71 or older 0.82 0.81 0.79 0.77 0.72 0.78 0.76 0.68 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 Targeting by proxy-means test One of the disadvantages of means testing is the difficulty to verify household income and assets especially in countries with a large informal sector, widespread subsistence agriculture and inflow of remittances. Although the current means test as applied for the MBPF performs relatively well in allocating the transfer to the poor households, coverage rates among the poor remain low (see 21 above). Recognizing the shortcomings of the current approach, it is worthwhile investigating the potential of proxy-means testing (PMT) for the identification of eligible households and allocation of the MBPF. With PMT, benefit eligibility is based on a set of observable indicators that are strongly correlated with poverty. Models can still include family income, but it is also common to include only indicators that are easy to observe. Since benefit receipt is not directly linked to the level of reported family income, work disincentives are of less importance. A more challenging aspect of PMT relates to information and communication. Since eligibility is based on scores, which are derived from econometric models, social offices may find it difficult explaining the process and potentially negative decisions to the population. The method also requires sufficient capacity both at central and local level for the regular calibration of the model and its correct implementation at the local level. The application process can be simplified thereby reducing the administrative burden for both client and officer. Documentation needs are usually lower under PMT compared to means tests. However, home visits may be necessary in order to verify the information provided by the applicants. Proxy-means tests use a limited number of indicators that are (i) easily observable, (ii) difficult to manipulate by households and (iii) strongly correlated with poverty. Together with the Social Policy Institute (SPI) a consumption model has been developed that can serve as the basis for the identification of the indicators to be included in the PMT. Based on discussions with the MSD, it was decided to include household income per capita as an additional indicator similar to the PMT model used in Georgia. In order to identify the relevant indicators, we estimate stepwise OLS regression models separately for urban and rural areas acknowledging differences in consumption patterns and asset ownership by area of residence. The calculated scores (coefficients) are subsequently used to predict a household’s economic welfare level and its eligibility for social transfers and services. The model used for the further analysis in this report is presented in annex 1. The estimated coefficients slightly deviate from the model developed by the SPI.9 In order to compare the predicted consumption levels with the original poverty status, we select the bottom 34% and bottom 6% of individuals based on predicted consumption. 71 percent of the absolute poor based on actual per capita consumption before MBPF are also identified as poor based on predicted consumption. The match is lower when considering extreme poverty. 58 percent of the extremely poor are identified likewise when using the predicted values.10 Table 17. Comparing actual and predicted poverty rates Absolute poverty by proxy Absolute Poverty not poor poor not poor 85.5 14.5 100 Poor 28.8 71.3 100 Extreme poverty by proxy Extreme Poverty Not poor poor not poor 97.1 2.9 100 poor 42.1 57.9 100 Note: actual poverty rates before MBPF; poverty by proxy based on predicted consumption. 9 It was not possible to exactly replicate the model developed by the SPI. The SPI model is also included in the annex for reference purposes. This had direct implications for the subsequent simulations. 10 Spearman rank correlation coefficient for predicted and actual (before MBPF) consumption per capita is 0.81 with p<0.000. 22 As with the regular means test, a cut-off threshold has to be determined for the identification of eligible households. Table 18 presents coverage and distribution rates for three different thresholds: the bottom 10 (p10), 15 (p15) and 20 (p20) percent of the population as identified by the PMT model. Since the PMT predicts benefit eligibility based on a set of indicators, there will always be exclusion and inclusion errors. Coverage of the poorest 20 percent of the population is highest with the most generous threshold (p20). Transfers would reach 67 percent of the poorest quintile. Setting the cut-off threshold at 10 percent would exclude 58 percent of the poorest 20 percent from receiving a transfer. Inclusion errors are highest with the most generous cut-off (bottom 20 percent), and lowest with the strictest definition of the cut-off threshold (bottom 15 percent). Table 18. Simulated coverage and distribution rates of different PMT cut-off thresholds, percentages Coverage Distribution of beneficiaries Quintile Bottom 10% Bottom 15% Bottom 20% Bottom 10% Bottom 15% Bottom 20% Q1 42.2 54.9 66.7 82.1 73.5 66.8 Q2 5.0 11.1 20.9 9.8 14.9 21.1 Q3 3.0 7.1 8.8 5.9 9.4 8.8 Q4 1.1 1.6 3.1 2.2 2.2 3.1 Q5 0.1 0.1 0.3 0.1 0.1 0.3 Total 10 15 20 100 100 100 Excl. error 57.8 45.2 33.3 Incl. error 18.0 26.5 33.2 Note: Quintiles are based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). Exclusion error = 100% - coverage Q1; Inclusion error = 100% - beneficiaries belonging to Q1. Source: KIHS 2010 Coverage rates of the poorest ten percent of the population (nationwide) differs across oblast. The difference in coverage rates is directly related to the regional living standards. The poorer the oblast, the larger the share of potential beneficiaries (table 19). Table 19. Simulated coverage of the poorest ten percent, by oblast, percentages p10 p15 p20 Issykul 34.3 49.4 58.6 Jalal-Abad 66.6 71.1 89.5 Naryn 61.3 77.7 82.7 Batken 19.6 45.2 51.4 Osh 74.4 84.4 91.3 Talas 39.7 40.3 46.7 Chui 47.5 51.3 62.3 Bishkek 0.0 16.6 16.6 Note: Decile is based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). It identifies the poorest ten percent at the country level. Source: KIHS 2010 For the subsequent simulation we estimate six different benefit scenarios for each of the three options. We further introduce a categorical targeting criteria similar to the current MBPF. Only households with children are eligible for a transfer.11 Flat benefits are either allocated per child or per household. Benefit amounts are derived from the available budget (low, medium, high) and the respective number of eligible children or households (table 20). 11 This has only minor implications for coverage rates as 99 percent of the households belonging to the bottom 20 percent identified by the PMT have children below the age of 16. 23 Table 20. Benefit scenarios for simulation of PMT Low Medium High Budget: 1,330 Budget: 2,000 Budget: 5,600 mio KGS mio KGS mio KGS Eligible children Monthly transfer per child below 16 (KGS) (#) P10: households with at least one child < 16 and 264,235 419 631 1,766 belonging to bottom 10 percent based on PMT P15: households with at least one child < 16 and 386,677 287 431 1,207 belonging to bottom 15 percent based on PMT P20: households with at least one child < 16 and 500,525 221 333 932 belonging to bottom 20 percent based on PMT Eligible Monthly transfer per household (KGS) households (#) P10: households with at least one child < 16 and 79,487 1,394 2,097 5,871 belonging to bottom 10 percent based on PMT P15: households with at least one child < 16 and 122,775 903 1,357 3,801 belonging to bottom 15 percent based on PMT P20: households with at least one child < 16 and 171,788 645 970 2,717 belonging to bottom 20 percent based on PMT Note: number of eligible children and households derived from KIHS 2010. Budget information from MSD. Table 21. Benefit distribution for different PMT options P10 and child <16 P15 and child <16 P20 and child <16 Quintile Benefit per child Benefit per hh Benefit per child Benefit per hh Benefit per child Benefit per hh Q1 82.3 80.4 74.7 72.7 68.9 65.0 Q2 10.5 11.1 14.7 15.1 19.9 22.1 Q3 5.1 5.6 8.5 9.4 8.1 8.9 Q4 2.0 2.8 2.1 2.7 2.9 3.7 Q5 0.1 0.1 0.1 0.1 0.2 0.4 Note: Quintiles are based on total household consumption per capita before MBPF (given marginal propensity of 33 percent). Source: KIHS 2010 Figure 8. Last resort social assistance programs: distribution of benefits to the poorest 20 percent (Source: WB) 24 Using a PMT to allocate transfers to poor families with children would further improve the targeting performance of the current MBPF, especially for the two lower cut-off thresholds (table 21). In the strictest case (P10), more than 80 percent of allocated transfers would be received by the poorest 20 percent of the population. This is a comparatively high score and would bring Kyrgyzstan among the top performers for last-resort social assistance schemes (figure 8). The distributional difference between flat benefits per child or per household is small, with a slight advantage for the former. In terms of poverty reduction effect, the impact is most noticeable when considering extreme poverty. Allocation transfers by PMT, even at the low budget scenario, would have a larger effect than the current MBPF (table 22). With respect to absolute poverty, the differences are small with respect to the poverty rate, but the reduction in the absolute poverty gap would increase. Table 22. Simulated impact on poverty for different PMT scenarios Absolute poverty Extreme poverty Rate (%) Gap (%) Rate (%) Gap (%) Before MBPF 33.93 7.55 5.63 0.84 After MBPF 33.68 7.43 5.34 0.79 P10: benefit per child low 33.77 7.23 4.36 0.62 med 33.75 7.06 3.99 0.54 high 33.25 6.21 2.07 0.29 P10: benefit per hh low 33.78 7.23 4.10 0.63 med 33.69 7.07 3.86 0.56 high 33.21 6.25 2.79 0.29 P15: benefit per child low 33.62 7.24 5.04 0.67 med 33.48 7.08 4.23 0.59 high 33.00 6.26 3.16 0.31 P15: benefit per hh low 33.76 7.24 5.28 0.67 med 33.54 7.09 3.99 0.60 high 32.92 6.29 2.73 0.33 P20: benefit per child low 33.73 7.24 5.15 0.70 med 33.43 7.09 4.77 0.63 high 32.81 6.29 2.97 0.38 P20: benefit per hh low 33.74 7.25 5.20 0.71 med 33.61 7.10 5.17 0.65 high 32.38 6.33 3.21 0.42 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 Table 23 presents extreme poverty rates and gaps for the different PMT scenarios for different age groups. Given the current budget (low scenarios), the extreme poverty risk for children could be decreased with the introduction of PMT. The effect is largest if eligibility for the transfer is limited to households belonging to the bottom 10 or 15 percent based on the PMT score. Overall, extreme poverty would be decline for almost all age groups. Table 23. Simulated impact on poverty for PMT scenarios, by age P10 (flat benefit per child) P15 (flat benefit per child) P20 (flat benefit per child) Before After MBPF MBPF low med high low med high low med high 25 Extreme poverty rate (%) below 6 10.06 9.44 7.54 6.87 2.41 9.10 7.33 5.46 9.46 8.69 5.39 6-15 6.12 5.54 4.67 4.47 2.46 4.96 4.43 3.25 5.21 4.59 2.95 16-20 4.51 4.35 3.23 2.76 2.00 3.98 3.13 2.13 4.15 3.98 2.07 21-40 6.87 6.58 5.31 4.85 2.37 6.38 5.20 3.88 6.46 6.01 3.59 41-60 2.90 2.90 2.51 2.30 1.33 2.78 2.46 1.94 2.78 2.67 1.94 61-70 2.89 2.89 2.89 2.04 1.25 2.89 2.87 2.01 2.45 2.45 1.34 71 or older 3.81 3.81 2.56 2.55 2.55 3.19 2.46 2.44 3.19 3.14 2.18 Extreme poverty gap (%) below 6 1.59 1.46 1.11 0.93 0.33 1.22 1.06 0.42 1.29 1.15 0.61 6-15 0.92 0.84 0.67 0.58 0.31 0.71 0.62 0.30 0.74 0.66 0.37 16-20 0.45 0.44 0.34 0.31 0.26 0.35 0.31 0.22 0.37 0.33 0.24 21-40 1.07 1.00 0.79 0.69 0.36 0.85 0.76 0.39 0.88 0.80 0.49 41-60 0.44 0.43 0.34 0.31 0.18 0.36 0.33 0.20 0.38 0.35 0.24 61-70 0.30 0.30 0.22 0.19 0.05 0.24 0.22 0.11 0.25 0.22 0.13 71 or older 0.82 0.81 0.72 0.70 0.55 0.73 0.69 0.55 0.74 0.71 0.59 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 Current methodology with increased GMI Given the relatively strong targeting performance of the current means test in terms of benefit distribution (see above), increasing the GMI may be sufficient to improve the effectiveness of the MBPF. Theoretically, by increasing the average benefit level we would expect higher coverage of the poorest quintile (reduction of the exclusion error) and a more pronounced effect on the reduction of extreme poverty. Simulating a means-tested transfer poses a number of modeling challenges. The accuracy of the means test depends on the truthfulness of the information provided by the applicants and the ability of the social office to verify the family income. Furthermore, the actual and expected benefit take-up rate needs to be taken into account. Raising the eligibility threshold, i.e. the GMI, is assumed to enlarge the pool of eligible households. We would expect that raising the GMI will increase the number of both applicants and eligible households. Ideally, these expectations are included in a model simulating the effect of such a policy change. Several studies have aimed at modeling take-up as such or changes in take-up due to a policy reform into the policy simulation. Methods include stochastic simulation models (Pudney et.al. 2006; Hancock et.al. 2003), qualitative choice models (Duclos 1997), or discrete choice models (Blundell et.al. 1988; Edmonds 2005; Duclos 1995; Younger 2003). Other studies modeled the potential impact of increased budget allocation by using marginal benefit incidence analysis or demand estimates (Younger 2003; Lanjouw & Ravallion 1999). The possibility of applying these models mainly depends on two factors (Blundell et.al. 1988): the quality of the data, such as underreporting, and the level of entitlement errors, such as measurement errors of entitlements and targeting errors. In the context of the current analysis, several factors prevent the simulation of a change in take-up following the increase of the GMI. First of all, the number of MBPF recipient households in the KIHS sample is very small. Secondly, family income as used for the administrative eligibility assessment cannot be directly estimated with KIHS data. For example, the application of land coefficients to 26 impute income from land cannot be replicated. Land coefficients vary per district and per land quality. The KIHS does not provide sufficient information to assign appropriate coefficients for land ownership. Furthermore, income from different sources may be underreported, either in the KIHS or during the application process. Thirdly, current MBPF receipts are underreported. This applies especially to the total annual benefit value. A detailed analysis of benefit receipt based on monthly data indicates that only 20 percent of the MBPF beneficiaries reported a transfer for each month in 2010. Another 20 percent reported a transfer during four months or less. Using monthly information increases the average monthly reported benefit considerably and brings its value closer to administrative information.12 Finally, the small number of observations combined with the observed targeting errors further reduces the use of the KIHS for advanced modeling. We tried to generate income variables that would better approximate the administrative income used for the means test (table 24). Each definition excludes an additional income source. Total household income excludes the value of food from own production. Income definition 2 excludes the MSB and MBPF. Income definition 4 further excludes local subsidies, but more importantly, income from private transfers, such as remittances. The latter is considered difficult to verify during the application process. The exclusion of private transfers reduces average monthly per capita income with almost 200 KGS. Income definition 5 excludes all social transfers and private transfers. The last column in table 24 estimates the percentage of the population with children having income below the GMI.13 Table 24. Average income levels for different income definitions Mean Min Max Below GMI KGS per month %* 1. Income per capita, w/o fik 2,026 5 23,544 3.9 2. Income per capita, w/o fik, msb, mbpf 2,015 5 23,544 5.1 3. Income pc, w/o fik, msb, mbpf,ls 2,010 5 23,544 5.1 4. Income pc, w/o fik, msb, mbpf, ls, private transfers 1,874 4 23,012 6.4 5. Income pc, w/o fik, pen,schol,si,msb,mbpf,ls, priv.trans. 1,555 2 19,923 14.2 *) Percentage of population living in households with children and income below GMI. GMI = 310 KGS per month. fik=food in kind; ls=local subsidies; schol=scholarships, si=social insurance benefits; pen=pensions. Source: KIHS 2010 Next, we model the probability of benefit take-up for households with children following Blundell et.al. (1988) and Edmonds (2005). The dependent variable is current MBPF receipt. Explanatory variables include income per capita without food in kind, MSB, MBPF, local subsidies and private transfers (definition 4), estimated entitlement, household characteristics and location indicators. Estimated entitlements are calculated by the difference between the GMI and income per capita times the number of eligible children. We estimate a binary choice model using a probit model, where benefit receipt (yi) is a function of different determinants xi : Table 25 presents the estimated results. While the level of estimated entitlements has no influence on the probability to receive a MBPF transfer, the level of per capita income matters. The higher the 12 Based on the monthly records from the KIHS 2010, the average MBPF is 221 KGS per child. This is reasonably close to the administrative average of 235 KGS for 2010. 13 Note that based on per capita household consumption, no household would fall below the GMI (see table 6). 27 income, the lower the probability of receiving a transfer. The same applies to the number of children in the household. All locational variables are relevant reflecting the fact that benefit receipt is more likely outside Bishkek and in rural and high mountain areas. Running the same model but with estimated entitlement based on a higher GMI level, returns the coefficient for entitlement significant as well. This could indicate that an increased GMI may lead to more applications. Table 25. Probability of benefit take-up Dep. Variable: MBPF receipt (1/0) Coefficient Std. Error z P>z Ln (entitlement) -0.022 0.035 -0.7 0.517 Ln (per capita income – definition 4) -0.286 0.065 -4.4 0.000 number of eligible children 0.291 0.050 5.8 0.000 household size 0.022 0.038 0.6 0.565 Age of hh head -0.058 0.019 -3.1 0.002 Age of hh head squared 0.000 0.000 2.6 0.009 Head has low education -0.112 0.158 -0.7 0.480 Female head 0.029 0.144 0.2 0.841 Hh has wage income 0.276 0.159 1.7 0.082 head is divorced 0.054 0.228 0.2 0.814 head is widowed 0.005 0.172 0.0 0.977 hh as agricultural equipment 0.010 0.113 0.1 0.926 hh has land 0.438 0.171 2.6 0.010 hh receives msb 0.517 0.310 1.7 0.095 Issykul 3.768 0.296 12.7 0.000 Jalal-Abad 4.392 0.289 15.2 0.000 Naryn 4.420 0.354 12.5 0.000 Batken 4.754 0.285 16.7 0.000 Osh 4.719 0.292 16.2 0.000 Talas 3.606 0.314 11.5 0.000 Chui 2.905 0.476 6.1 0.000 rural area 0.233 0.088 2.6 0.008 high mountain area 0.394 0.211 1.9 0.062 constant -3.109 0.834 -3.7 0.000 pseudo R2 0.2827 n 3332 Note: includes only households with children. Robust standard errors. Source: KIHS 2010 In a next step we estimate a 2-stage selection model (Type-2 tobit model), where we estimate the probability of receiving a benefit in the first stage and the entitlement level in the second stage, but only for recipient households (results in annex 2). The model finds that the income level is negatively correlated both with the probability of receiving the benefit and with the size of estimated benefit entitlements for recipient households. However, the number of eligible children is no longer significant in the second stage. Finally, we apply a switching regression model where the average monthly MBPF receipt is regressed on (i) income, the level of the GMI, the number of eligible children and all other household and location characteristics as above, and (ii) a series of interaction terms for all previous variables with a dummy variable for theoretical benefit entitlement. The purpose of the model is to predict benefits assuming a higher GMI. Due to a negative coefficient for the level of the GMI in the model, contrary to what is expected, the predictions cannot be made. 28 Since none of the models allowed to predict future entitlements given a higher GMI, we have to base the subsequent simulations on the current situation. This means that we assume that no additional households will apply once the GMI is increased. Whether this leads to an under- or overestimate of the simulation results is difficult to say, as new applications could turn out to be wrongly included or excluded. For the simulation we assume that households receive the MBPF during the whole year, i.e. for 12 months. This means that we first replace the annual values in the survey by the reported monthly values multiplied by 12 (also see above). This will present the low scenario. The increase in GMI is based on forecasts from the MSD. The medium scenario foresees an increase of the GMI to 530 KGS per month with an estimated average MBPF of 446 KGS. In the high scenario, the GMI would be raised to 1050 KGS per month, resulting in an average MBPF of 698 KGS. For the simulation we increase the current MBPF benefits with the respective percentage increase for all beneficiary households. Table 26. Simulation scenarios for increasing the GMI Low medium high GMI per month (KGS) 310 530 1050 Average MBPF per month (KGS) 235 446 698 Increase with respect to low scenario 90% 197% For comparative purposes, table 27 shows the coverage of the poorest ten percent by the current MBPF. No one lives in Chui oblast or Bishkek that belongs to the poorest ten percent in the country and receives an MBPF. The highest coverage rates are observed in Osh, Naryn and Batken. Table 27. Coverage of the poorest ten percent, by oblast, percentages Current MBPF Issykul 12.1 Jalal-Abad 25.6 Naryn 35.7 Batken 33.9 Osh 56.5 Talas 7.0 Chui 0 Bishkek 0 Source: KIHS 2010 Table 28 and 29 show the simulated impact of different GMI levels on poverty. Although the low scenario is based on the actual GMI in 2010 (310 KGS per month), the simulation shows slightly lower poverty rates for the low scenario compared to the current situation after MBPF. This is the result of assigning beneficiary households the reported monthly MBPF for the full year. Increasing the GMI has only limited effects on poverty. This is partly due to the effect that the simulation assumed no change in coverage of the poorest households. Furthermore, the number of observations for eligible households as found in the KIHS is small. Therefore, the impact on poverty remains limited. 29 Table 28. Simulated impact on poverty for different GMI scenarios Absolute poverty Extreme poverty Rate (%) Gap (%) Rate (%) Gap (%) Before MBPF 33.93 7.55 5.63 0.84 After MBPF 33.68 7.43 5.34 0.79 GMI low 33.63 7.39 5.34 0.78 med 33.54 7.25 5.21 0.74 high 33.33 7.08 5.06 0.70 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 Table 29. Simulated impact on poverty for different GMI scenarios, by age GMI Before MBPF After MBPF low med high Extreme poverty rate (%) below 6 10.06 9.44 9.44 9.23 8.87 6-15 6.12 5.54 5.54 5.30 5.02 16-20 4.51 4.35 4.35 4.35 4.35 21-40 6.87 6.58 6.58 6.37 6.22 41-60 2.90 2.90 2.90 2.89 2.86 61-70 2.89 2.89 2.89 2.89 2.89 71 or older 3.81 3.81 3.81 3.76 3.76 Extreme poverty gap (%) below 6 1.59 1.46 1.44 1.32 1.20 6-15 0.92 0.84 0.83 0.78 0.73 16-20 0.45 0.44 0.44 0.44 0.43 21-40 1.07 1.00 0.99 0.93 0.88 41-60 0.44 0.43 0.42 0.41 0.40 61-70 0.30 0.30 0.30 0.30 0.30 71 or older 0.82 0.81 0.81 0.80 0.79 Note: Poverty rates before and after transfers take into account substitution effects (marginal propensity of 33 percent). Source: KIHS 2010 Cost efficiency of alternative targeting options On first sight, the different options simulated above do not differ considerably with respect to their efficiency and effectiveness. For a better comparison we calculate the costs of reducing the poverty gap with one KGS for each option. Note that we only include the costs of transfers. Administrative costs are not included. The administrative costs for categorical targeting are expected to be at the same level as for the current means test, or lower. Although the abolition of the means test will free- up time for the local staff, the increased case load due to higher number of beneficiaries has to be taken into account. Moving to a PMT is especially costly during start-up. A completely new system has to be introduced requiring extensive capacity building of staff and communication and information activities. Once the new system is in place, administrative costs might be at a similar level as with the current means test. Figure 9 compares the benefit cost efficiency of the various 30 options. Compared to the current MBPF, only PMT targeting would be more cost efficient in terms of absolute poverty reduction. With respect to the reduction of the extreme poverty gap (see figure 10 in annex 3), this conclusion holds as well, especially if comparing medium and high scenarios with medium and high GMI scenarios. Interestingly, the strictest categorical targeting option (e2) at medium level with benefits allocated per child is more cost-efficient than increasing the GMI to medium level. Figure 9. Cost efficiency of simulated options – reduction of absolute poverty gap (without administrative costs) high child hh child hh child p20 hh GMI med low high med low high p10 p10 p15 p15 p20 med low high med low high med low high med low high med low e6 hh high med low high ent child e2 hh child e6 med low high med low high curr e2 med low MBPF 0.00 1.00 2.00 3.00 4.00 5.00 6.00 7.00 Cost per 1 KGS reduction of absolute poverty gap Conclusion The objective of this report was to analyze and discuss the linkages between different minimum living standards currently used in the Kyrgyz Republic and to analyze the potential of changing the targeting method and/or benefit level for the MBPF. The different minimum standards all serve different purposes. The national poverty lines are first and foremost meant to analyze the observed living standards and measure policy outcomes. The MSL also serves as a standard to measure the living standard of the population. But it is a normative standards, which is also its main difference to the poverty lines, which are derived empirically. The GMI, on the other hand, is a purely budget-driven minimum standard. The difference in levels and population shares below the level can be largely explained by differences in indicators used to (i) measure poverty, and (ii) assess benefit eligibility. The former is based on household consumption, 31 while the latter uses household income. Moreover, administrative family income used for the eligibility assessment cannot be fully replicated with the KIHS data. The second part of the report analyzed the potential for alternative targeting methods and benefit levels using static microsimulation. Categorical targeting, proxy-means testing and the raising the current GMI were analyzed by looking at potential exclusion and inclusion errors as well as the estimated impact on poverty. Finally, all simulated options were compared based on their cost effectiveness in reducing the poverty gap. In order to reduce exclusion errors, a move to categorical or PMT targeting might be considered. Raising the GMI given the current means test seems to be less cost efficient in reducing poverty compared to a system based on PMT or even some strict categorical definitions. However, moving to a completely new targeting system may entail considerable costs, especially during start-up. Furthermore, the actual implementation may lead to different results as simulated in this report. Both, the models for categorical and PMT targeting assumed perfect implementation. This is highly unlikely as some households may not take-up the benefits. 32 References Blundell, R., Fry, V. and I. Walker (1988), Modelling the Take-up of Means-Tested Benefits: The Case of Housing Benefits in the United Kingdom, in The Economic Journal, 98(390), pp. 58-74. Duclos, J. (1997), Estimating and Testing a Model of Welfare Participation: the Case of Supplementary Benefits in Britain, in Economica, 64, pp. 81-100. Duclos, J. (1995), Modelling the take-up of state support, in Journal of Public Economics, 58, pp. 391- 415. Edmonds, E.V. (2005), Targeting Child Benefits in a Transition Economy, in Economics of Transition, 13(1), pp. 187-210. Government of the Kyrgyz Repulic (2011), Methodology of Poverty Line Estimation, Government Resolution of March 25, 2011, no. 115. Hancock, R., Pudney, S. and H. Sutherland (2003), Using Econometric Models of Benefit Take-up by British Pensioners in Microsimulation Models, Conference paper. https://guard.canberra.edu.au/natsem/conference2003/papers/pdf/hancock_ruth-paper2.pdf Lanjouw, P. and M. Ravallion (1999), Benefit Incidence, Public Spending Reform, and the Timing of Program Capture, in The World Bank Economic Review, 13(2), pp. 257-273. Pudney, S., Hancock, R. and H. Sutherland (2006), Simulating the Reform of Means-tested Benefits with Endogenous Take-up and Claim Costs, in Oxford Bulletin of Economics and Statistics, 68(2), pp. 135-166. Tsirunyan, S. (2005), Methodology of Calculation of Poverty Line Based on Kyrgyz Integrated Household Surveys (KIHS) 2003, Report for the World Bank. Tsirunyan, S. (2012), Methodology of Calculation of Poverty Line Based on Kyrgyz Integrated Household Surveys (KIHS) 2011, Report for the World Bank. Younger, S. (2003), Benefits on the Margin: Observations on Marginal Benefit Incidence, in The World Bank Economic Review, 17(1), pp. 89-106. 33 Annex 1: Model for proxy-means test We use a weigthed stepwise OLS regression model in order to find the most powerful indicators that are highly correlated with poverty among a long list of indicators. The dependent variable is the natural logarithm of per capita consumption before MBPF transfers (66 percent substitution assumed). Normalizing the absolute consumption values by taking the logarithm reduces the estimation bias caused by the skewed distribution of per capita consumption. The estimated model can be written as follows: Ln(yi) = α + βXi + εi where yi is pre-transfer household consumption per capita, α a constant, β a vector of coefficients, Xi a vector of household indicators, and εi the error term. We estimate the models separately for urban and rural areas accounting for the different living conditions. The selection criteria for inclusion of the indicators in the stepwise regression was set at p < 0.10. Table 30 presents the estimation results. Table 30. Proxy-means indicators for household consumption per capita* Urban model Rural model Coeff. St. error P>t Coeff. St. error P>t ln (pc income) 0.308 0.010 0.000 0.099 0.008 0.000 household size -0.071 0.005 0.000 -0.103 0.005 0.000 number of cows 0.054 0.015 0.000 0.076 0.008 0.000 hh with 2 children -0.031 0.011 0.005 -0.136 0.023 0.000 hh has washing machine 0.038 0.011 0.000 0.090 0.014 0.000 number of goats -0.010 0.005 0.048 0.004 0.002 0.012 hh with 5 or more children -0.310 0.035 0.000 hh uses oven or fireplace for cooking -0.025 0.014 0.073 hh head is female -0.034 0.013 0.010 hh head is widowed -0.027 0.016 0.097 hh has mobile phone 0.133 0.013 0.000 0.092 0.016 0.000 hh head was never married 0.086 0.032 0.007 hh members with high education 0.029 0.006 0.000 0.059 0.009 0.000 Bathroom outside the house -0.071 0.013 0.000 hh has garage 0.072 0.021 0.001 0.098 0.025 0.000 number of poultry 0.007 0.002 0.000 0.003 0.001 0.005 hh members with less than sec school 0.013 0.007 0.056 hh head is unemployed -0.096 0.022 0.000 number fo pensioners -0.063 0.007 0.000 -0.016 0.009 0.094 hh has computer 0.088 0.020 0.000 0.162 0.038 0.000 has has satellite dish 0.078 0.023 0.001 0.073 0.030 0.015 Issykul -0.094 0.021 0.000 -0.083 0.025 0.001 Jalal-Abad -0.181 0.016 0.000 -0.145 0.017 0.000 Naryn -0.248 0.036 0.000 -0.262 0.029 0.000 Batken -0.088 0.025 0.001 Osh -0.180 0.016 0.000 Talas -0.105 0.038 0.006 0.091 0.030 0.003 Chui -0.099 0.020 0.000 living rooms per person 0.134 0.012 0.000 34 additional house 0.166 0.040 0.000 hh has car 0.062 0.013 0.000 0.087 0.018 0.000 number of horses 0.323 0.080 0.000 0.023 0.012 0.051 hh has one child -0.065 0.022 0.003 hh has three children -0.104 0.027 0.000 hh has four children -0.167 0.032 0.000 hh members with own business 0.037 0.015 0.011 hh members with sec prof edu 0.032 0.011 0.004 number of old-age pensioners 0.053 0.019 0.006 ln (land) 0.017 0.004 0.000 walls not made of brick, concrete slabs or wood -0.052 0.017 0.002 central heating 0.168 0.066 0.011 water supply 0.113 0.026 0.000 Constant 5.627 0.089 0.000 7.177 0.073 0.000 Number of observations 2957 1945 Adjusted R2 0.704 0.579 * Dependent variable: natural logarithm of household consumption per capita before MBPF (marginal propensity of 33%). Source: own calculations based on KIHBS 2005. The estimated model coefficients are the scores for the proxy-means test. For each household, a score is calculated multiplying the indicators with the coefficients and returning the logarithmic values back to absolute values. The variables included in the model above were provided by the Social Policy Institute (SPI) that is currently supporting the MSD. The aim was to use the same PMT model for the simulations in this report. However, it was not possible to exactly replicate the SPI model even though the datasets were shared. Below are the models developed by the SPI for comparison. Urban areas: Model Summary Adjusted Std. Error of Model R R Square R Square the Est imat e 1 ,825a ,681 ,678 ,27442 a. Predictors: (Constant), n2_horse, obl_1, pc, unempl_ hh, nev ermarried, business, det 1n, ad_housing, garage, obl_2, n_Higher_educ_sum, pens_sum, obl_ 4, n2_poultry , n_rooms, antena, obl_5, wash12, n_ cow, obl_6, car, mobile, det 2n, WC_out, obl_3, inc_pc_ out_s_ln, HH size 35 a Coeffici ents Unstandardized Standardized Coef f icients Coef f icients Model B Std. Error Beta t Sig. 1 (Constant) 5,671 ,091 62,061 ,000 inc_pc_out_s_ln ,314 ,011 ,405 29,003 ,000 HH size -,095 ,005 -,337 -19,393 ,000 det1n -,033 ,013 -,029 -2,527 ,012 det2n -,036 ,015 -,029 -2,433 ,015 nev ermarried ,089 ,028 ,035 3,207 ,001 unempl_hh -,067 ,022 -,032 -2,997 ,003 pens_sum -,054 ,007 -,084 -7,327 ,000 n_Higher_educ_sum ,034 ,006 ,059 5,165 ,000 business ,111 ,041 ,029 2,705 ,007 obl_1 -,035 ,018 -,025 -1,961 ,050 obl_2 -,146 ,018 -,105 -8,275 ,000 obl_3 -,217 ,021 -,128 -10,281 ,000 obl_4 -,082 ,021 -,048 -3,870 ,000 obl_5 -,115 ,017 -,081 -6,600 ,000 obl_6 -,068 ,021 -,040 -3,183 ,001 n_rooms ,091 ,009 ,147 9,851 ,000 WC_out -,068 ,013 -,068 -5,268 ,000 garage ,045 ,018 ,029 2,511 ,012 ad_housing ,099 ,036 ,029 2,756 ,006 car ,074 ,015 ,058 4,955 ,000 pc ,123 ,024 ,055 5,041 ,000 antena ,092 ,020 ,051 4,576 ,000 mobile ,115 ,013 ,109 9,096 ,000 wash12 ,047 ,011 ,048 4,161 ,000 n_cow ,078 ,017 ,051 4,496 ,000 n2_poult ry ,008 ,002 ,056 4,977 ,000 n2_horse ,118 ,051 ,025 2,297 ,022 a. Dependent Variable: pccd_M_out_s_ln Rural areas: Model Summary Adjusted Std. Error of Model R R Square R Square the Est imat e 1 ,767a ,588 ,582 ,30422 a. Predictors: (Constant), agrequipm, det2n, gaz_oper, n_SecProf _educ_sum, unempl_hh, n_Higher_educ_ sum, obl_1, antena, mobile, pc, det4n_, n2_poultry , n2_goat, obl_2, garage, pens_sum, wash12, det3n, obl_6, car, wat er_oper, n2_horse, S_plot_ln, n_cow, inc_pc_out_s_ln, obl_3, det1n, HH size 36 a Coeffi ci ents Unstandardized St andardized Coef f icients Coef f icients Model B St d. Error Beta t Sig. 1 (Constant) 7,416 ,068 108,608 ,000 inc_pc_out _s_ln ,075 ,007 ,186 10,162 ,000 HH size -,138 ,006 -,537 -21,396 ,000 det1n -,063 ,022 -,057 -2,878 ,004 det2n -,126 ,025 -,113 -5,030 ,000 det3n -,085 ,030 -,065 -2,814 ,005 det4n_ -,082 ,037 -,054 -2,215 ,027 n_Higher_educ_sum ,050 ,011 ,073 4,693 ,000 n_SecProf _educ_sum ,038 ,012 ,046 3,104 ,002 unempl_hh -,055 ,029 -,029 -1,896 ,058 pens_sum -,019 ,010 -,031 -1,810 ,070 obl_1 -,075 ,023 -,055 -3,235 ,001 obl_2 -,135 ,023 -,095 -5,859 ,000 obl_3 -,221 ,025 -,161 -8,855 ,000 obl_6 ,126 ,024 ,092 5,346 ,000 water_oper ,098 ,029 ,057 3,337 ,001 gaz_oper ,131 ,040 ,052 3,326 ,001 garage ,124 ,025 ,079 4,887 ,000 car ,108 ,019 ,090 5,561 ,000 pc ,232 ,052 ,069 4,468 ,000 antena ,084 ,030 ,044 2,809 ,005 mobile ,065 ,016 ,064 4,068 ,000 wash12 ,066 ,015 ,069 4,328 ,000 S_plot_ln ,023 ,004 ,100 5,752 ,000 n_cow ,084 ,008 ,185 10,475 ,000 n2_goat ,005 ,001 ,050 3,015 ,003 n2_poult ry ,005 ,001 ,063 4,027 ,000 n2_horse ,017 ,008 ,037 2,199 ,028 agrequipm ,040 ,018 ,036 2,202 ,028 a. Dependent Variable: pccd_M_out_s_ln 37 Annex 2: Alternative models for modeling benefit take-up 1. Type-2 tobit model Probit regression Number of obs = 3332 LR chi2(22) = 544.35 Prob > chi2 = 0.0000 Log likelihood = -691.53344 Pseudo R2 = 0.2824 ------------------------------------------------------------------------------ d_umb | Coef. Std. Err. z P>|z| [95% Conf. Interval] -------------+---------------------------------------------------------------- ln_pcfy3 | -.2523902 .0391052 -6.45 0.000 -.3290349 -.1757454 n_gmi | .2911457 .0536265 5.43 0.000 .1860398 .3962516 hsize | .0245057 .0441402 0.56 0.579 -.0620074 .1110188 age | -.0581834 .0191806 -3.03 0.002 -.0957767 -.02059 age2 | .0004761 .0001854 2.57 0.010 .0001127 .0008395 d_lowedu | -.1128453 .1541475 -0.73 0.464 -.4149688 .1892783 f_head | .0266887 .1402354 0.19 0.849 -.2481676 .301545 d_incwage | .2741455 .1453254 1.89 0.059 -.0106872 .5589781 d_divorced | .0649842 .2359928 0.28 0.783 -.3975533 .5275216 d_widowed | .0137846 .1656947 0.08 0.934 -.3109711 .3385403 d_agequip | .0070303 .1115631 0.06 0.950 -.2116294 .22569 d_land | .4400522 .1685047 2.61 0.009 .1097892 .7703153 d_msb | .5153132 .2732005 1.89 0.059 -.0201499 1.050776 obl1 | 3.766478 113.7353 0.03 0.974 -219.1507 226.6837 obl2 | 4.388291 113.7353 0.04 0.969 -218.5288 227.3054 obl3 | 4.411022 113.7355 0.04 0.969 -218.5064 227.3285 obl4 | 4.752949 113.7353 0.04 0.967 -218.1642 227.6701 obl5 | 4.716819 113.7353 0.04 0.967 -218.2003 227.6339 obl6 | 3.606594 113.7354 0.03 0.975 -219.3106 226.5238 obl7 | 2.891972 113.7358 0.03 0.980 -220.0261 225.8101 u_r2 | .2385113 .0900523 2.65 0.008 .062012 .4150106 d_mount | .4031583 .2065906 1.95 0.051 -.001752 .8080685 _cons | -3.440484 113.7369 -0.03 0.976 -226.3607 219.4797 ------------------------------------------------------------------------------ Source | SS df MS Number of obs = 281 -------------+------------------------------ F( 21, 259) = 42.43 Model | 1928.54596 21 91.8355217 Prob > F = 0.0000 Residual | 560.533544 259 2.16422218 R-squared = 0.7748 -------------+------------------------------ Adj R-squared = 0.7565 Total | 2489.0795 280 8.88956965 Root MSE = 1.4711 ------------------------------------------------------------------------------ ln_ent | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- ln_pcfy3 | -1.623377 .0810701 -20.02 0.000 -1.783018 -1.463737 n_gmi | .0302432 .1436253 0.21 0.833 -.2525787 .3130651 hsize | -.2090389 .1320273 -1.58 0.115 -.4690224 .0509447 age | .0168692 .0492159 0.34 0.732 -.080045 .1137834 age2 | -.0001331 .0004763 -0.28 0.780 -.0010711 .0008049 d_lowedu | -.2012735 .3879065 -0.52 0.604 -.9651255 .5625786 f_head | -.0541882 .3632687 -0.15 0.882 -.7695245 .6611481 d_incwage | .2316895 .3135998 0.74 0.461 -.3858405 .8492195 d_divorced | -.6532749 .6556171 -1.00 0.320 -1.944293 .6377437 d_widowed | -.4678354 .4482064 -1.04 0.298 -1.350428 .4147572 d_agequip | .4622653 .254635 1.82 0.071 -.0391531 .9636838 d_land | -.1095685 .5258476 -0.21 0.835 -1.14505 .9259125 d_msb | .2642892 .5980016 0.44 0.659 -.9132749 1.441853 obl1 | -1.745937 1.621487 -1.08 0.283 -4.938912 1.447039 obl2 | -1.266667 1.602454 -0.79 0.430 -4.422165 1.888831 obl3 | -1.006365 1.638839 -0.61 0.540 -4.233511 2.220781 obl4 | -1.805582 1.581693 -1.14 0.255 -4.920198 1.309034 obl5 | -1.620636 1.587865 -1.02 0.308 -4.747405 1.506132 obl6 | -1.951749 1.639767 -1.19 0.235 -5.180723 1.277224 obl7 | (omitted) u_r2 | -.3673468 .231282 -1.59 0.113 -.8227793 .0880857 d_mount | -.766498 .4372629 -1.75 0.081 -1.627541 .0945451 _cons | 18.10215 2.278335 7.95 0.000 13.61574 22.58857 38 2. Switching regression model Linear regression Number of obs = 3332 F( 45, 3286) = 12.98 Prob > F = 0.0000 R-squared = 0.2158 Root MSE = 1.5934 ------------------------------------------------------------------------------ | Robust lnmumb | Coef. Std. Err. t P>|t| [95% Conf. Interval] -------------+---------------------------------------------------------------- dlnpcfy3 | .4263656 .1608561 2.65 0.008 .1109773 .741754 dvgmi | -.0172137 .007016 -2.45 0.014 -.0309698 -.0034576 dn_gmi | .0358344 .2310168 0.16 0.877 -.4171171 .4887859 dhsize | .2924787 .1957706 1.49 0.135 -.0913659 .6763233 dage | .0433226 .0914671 0.47 0.636 -.1360157 .2226608 dage2 | -.0005977 .0010098 -0.59 0.554 -.0025776 .0013821 ddlowedu | -.3812912 .4367076 -0.87 0.383 -1.237538 .4749554 df_head | -.1816839 .4808505 -0.38 0.706 -1.124481 .7611131 ddincwage | -.851428 .5142836 -1.66 0.098 -1.859777 .1569207 dddivorced | -.3664674 .9639066 -0.38 0.704 -2.256386 1.523451 ddwidowed | .2346539 .678201 0.35 0.729 -1.095085 1.564393 ddagequip | -.0790017 .4408598 -0.18 0.858 -.9433895 .7853861 ddland | 1.466775 .6690525 2.19 0.028 .1549734 2.778577 ddmsb | 2.652455 .9731739 2.73 0.006 .7443669 4.560544 dobl1 | -.1202268 .7734345 -0.16 0.876 -1.636689 1.396236 dobl2 | 2.27545 .9352396 2.43 0.015 .4417391 4.109162 dobl3 | .0167864 1.488873 0.01 0.991 -2.902427 2.936 dobl4 | 2.364094 1.060536 2.23 0.026 .2847164 4.443472 dobl5 | 1.556326 .8677961 1.79 0.073 -.1451495 3.257802 dobl6 | -1.301817 .7632963 -1.71 0.088 -2.798401 .1947676 dobl7 | -1.006736 .8230802 -1.22 0.221 -2.620538 .6070658 durbrur | -1.137448 .5774693 -1.97 0.049 -2.269684 -.0052116 ddmount | .8979756 1.271255 0.71 0.480 -1.594556 3.390508 ln_pcfy3 | -.3578935 .0622737 -5.75 0.000 -.4799926 -.2357944 v_gmi | (omitted) n_gmi | .2782909 .0466295 5.97 0.000 .186865 .3697168 hsize | -.029986 .0285043 -1.05 0.293 -.0858739 .025902 age | -.0425962 .0161371 -2.64 0.008 -.0742359 -.0109565 age2 | .0003695 .0001546 2.39 0.017 .0000663 .0006727 d_lowedu | -.0875606 .1207588 -0.73 0.468 -.3243306 .1492095 f_head | .1292146 .1021268 1.27 0.206 -.0710241 .3294532 d_incwage | .1256947 .1439987 0.87 0.383 -.1566416 .4080309 d_divorced | -.0601917 .1120203 -0.54 0.591 -.2798283 .159445 d_widowed | -.1547938 .1158749 -1.34 0.182 -.3819881 .0724005 d_agequip | -.1268566 .11365 -1.12 0.264 -.3496887 .0959754 d_land | .1242595 .0649462 1.91 0.056 -.0030796 .2515986 d_msb | .1034007 .2476807 0.42 0.676 -.3822234 .5890249 obl1 | -.1864246 .0645318 -2.89 0.004 -.3129512 -.059898 obl2 | .0936988 .069646 1.35 0.179 -.0428552 .2302528 obl3 | .2328626 .3588665 0.65 0.516 -.4707619 .9364872 obl4 | .5803277 .1109869 5.23 0.000 .3627173 .7979381 obl5 | .4484145 .0905632 4.95 0.000 .2708484 .6259805 obl6 | -.1613494 .0802073 -2.01 0.044 -.3186107 -.004088 obl7 | -.2442419 .0611296 -4.00 0.000 -.364098 -.1243859 u_r2 | .2343226 .0745059 3.15 0.002 .0882399 .3804052 d_mount | .3763997 .3427784 1.10 0.272 -.2956811 1.048481 _cons | 4.17438 .7729143 5.40 0.000 2.658937 5.689822 --------------------------------------------------------------------------------------- 39 Annex 3: Cost efficiency extreme poverty Figure 10. Cost efficiency of simulated options – reduction of extreme poverty gap (without administrative costs) high t child e2 hh child e6 hh child hh child hh child p20 hh GMI med low high med low high p10 p10 p15 p15 p20 med low high med low high med low high med low high med low high med low high e6 med low high med low high ren e2 med low cur MBPF 0.00 5.00 10.00 15.00 20.00 25.00 30.00 35.00 40.00 Costs per 1 KGS reduction of extreme poverty gap 40