Report No. 39737-XK Kosovo Poverty Assessment (In Two Volumes) Volume II: Estimating Trends from Non-comparable Data October 3, 2007 Poverty Reduction and Economic Management Unit Europe and Central Asia Region Document of the World Bank VOLUME I1 TABLE OF CONTENTS CHAPTER 1: HOUSEHOLD BUDGET SURVEY (HBS) AND POVERTYMONITORING IN Kosovo ..................................................................................................................... A. There are Problemsof Data Comparability ............................................................. 7 8 (a) Problem # 1:Diary versus Recall.................................................................... 8 (b) Problem #2: Survey Design-Redefinition of Consumption Items................8 B(c)Sample Weights Introduce Additional Uncertainty................................................. 9 9 C 10 D Likely Consequences:Poverty Estimates.............................................................. ... Problem #3: Survey Design - LSMS versus HBS........................................... Likely Consequences:Consumption..................................................................... 12 CHAPTER 2: A. Post-Stratification .................................................................................................. POVERTY-ALTERNATIVEESTIMATES ...................................................... 15 17 B. Compare Only 2003 and 2005............................................................................... 19 D. Compare all the Years ........................................................................................... C. Comparable Consumption Aggregate Methodology............................................. 20 E. Comparison of Poverty Figures from the LSMS and HBS................................... 22 23 CHAPTER 3: A Recommendations................................................................................................ . CONCLUSIONS AND RECOMMENDATIONS................................................... 25 -25 ANNEXA: TABLES AND FIGURES ..................................................................................... 27 ANNEXB: RESULTS USING DIFFERENT SURVEY YEARDEFINITION................................ 33 3 ListofTables Table 1.1: PopulationSize by Survey Wave andYear ................................................................. 10 Table 1.2: Summary of Survey ConstraintsandTheir Effectson PovertyEstimates..................12 Table 1.3: PovertyHeadcountbyLocationandEthnicareas, usingPA05 methodology............13 Table 2.1: Overviewof the Resultsof Methodologies for ComparablePovertyEstimates.........14 Table 1.4: PovertyHeadcountby HouseholdHeadEthnicity...................................................... Table 2.2: Summary of Poverty Estimatesfromthe MethodologiesUsed................................... 17 17 Table 2.3: PovertyRateswith CurrentWeights andReweighted ................................................ Table 2.4: Samplingprocedurefor the Bosniaand Herzegovina'sHouseholdBudget Survey .... 18 19 Table 2.5: PovertyRateswith the PA05 and ComparableCA methodologies............................ 20 Table 2.6: RobustPovertyLinesBasedon ConsistentFoodItems .............................................. Table 2.7: PovertyRatesusingthe AbbreviatedConsumptionBundleMethodology .................21 21 Table A.1: ComparisonofPreviousMethodologies .................................................................... 27 Table A.2: Survey Comparison Table A.3: PercentChangesinMainAggregates from Survey to Survey Comparison...............29 .................................................................................................... 30 Table A.4: AlternativeConsumptionAggregate DefinitionsandPovertyRates .......................... 30 Table A.6: Definitionof ConsumptionAggregates for the DifferentMethodologies..................-32 Table A.5: Consistently Asked Questionsover the Four Surveys ................................................ 31 Table A.7: Poverty Linesin DifferentMethodologies................................................................. 32 Table B.1: IntroductionofNewQuestionnaires........................................................................... 33 Table B.2: PovertyStatistics usingPA05 Methodology Table B.3: PovertyRatesUsingPA05 Methodology................................................................... .............................................................. 33 Table B.4: DetailedPoverty Diagnostics with RevisedConsumptionAggregate.,...................... 34 Table B.5: PovertyRatesUsingAlternativeConsumptionandPoverty LineMethodologies .....34 35 List of Boxes Box 2.1: BosniaandHerzegovinaHBS: Exampleof Samplingwithout a Census..................... Box 2.2: Analysis ofChanges...................................................................................................... 19 22 ListofFigures Figure 1.1: Total Populationin Millions and HouseholdSize ....................................................... Figure 1.2: Average MonthlyHouseholdConsumption, inNominalprices................................ 9 Figure 1.3: PovertyRateEstimatesandthe EffectofChanges inthe Questionnaire...................10 13 the Population................................................................................................................................ Figure2.1: Cumulativeand DensityDistributionofConsumption for the Bottom50 percentileof 18 4 ACKNOWLEDGEMENTS This report is a joint production of Statistical Office of Kosovo staff in the Household Budget Survey unit comprising Bashkim Bellaqa, Bekim Canoli, and Emina Deliu and World Bank technical team comprising Andrew Dabalen and Anna Gueorguieva, supported by Sasun Tsirunyan and Shpend Ahmeti. The report has benefited from the support of UK's Department for International Development which has generously funded the Trust Fund to support the capacity buildingand analyticactivitiesofthe WesternBalkanProgrammaticPovertywork. The reportwould not have beenpossible without the very close involvementand support ofthe Social Statistics Department ofthe StatisticalOfficeofKosovo. The team graciously acknowledgesthe analyticwork of the IMF (Macro statistics), EAR and Ministry of Agricultureof the PISG, SOK andVllaznim Bytyqi(Migration). The team has benefited from the comments and guidance of Peter Lanjouw (Peer Reviewer), Pierella Paci (Peer Reviewer), Asad Alam, Ardo Hansson, ElisabethHuybens, Felix Martin, Kanthan Shankar, Julian Lampietti, Ruslan Yemtsov, Kinnon Scott, Gero Carletto, Marcus Goldstein, Gabriel Demombynes, Juan Munoz, and Johan Mistiaen for excellent commentsand suggestions. The productionofthis reportbenefited enormouslyfrom the excellenteditingskills of SusanaPadilla. 5 6 CHAPTER 1: HOUSEHOLDBUDGET SURVEY(HBS)AND POVERTYMONITORINGINKOSOVO Since 2002, Kosovo has conducted annual Household Budget Surveys (HBS). At first glance, availability of annual cross-sections of detailed collection of household consumption expenditure data should suggest that one should be able to track poverty and inequality over time. However, examining changes in poverty and inequality over time in Kosovo poses several challenges. The main problem is data comparability because of (9 changes in survey design and (ii) large sampling errors. First, a wide variety of experience in other countries has shown that even small changes in the way expenditure/consumption or income data is collected can have a substantial impact on poverty estimates. These experiences have documented that differences in the poverty estimates over time could be driven by changes in survey design rather than by a real change in household welfare. The survey sampling weights, on the other hand, compound theproblem as they introduce an unquantijable bias or sampling error. The sampling was based on an outdated population frame and with limited survey supervision. In this note, we apply several methods to construct poverty estimates that are consistent over time. First, we make an attempt to construct a comparable consumption by aggregating items that were defined uniformly andfocusing only on the years where the questionnairedid not change. Second, we use an adjustmentprocedure that relies on afew variables whose definition has not changed over time to update the distribution of thepoor over time. The resultsfrom these various methods show that during the period from 2002 to 2006, poverty was high, at around 45 percent, and that there is no evidence of a sustainedimprovementin the welfare of households in Kosovo. The recommendationsfor data collection for poverty monitoring coming from this research are to,first, maintain consistency in the survey questionnaire, second, to conduct a population census, and, third, to emphasize better survey administration and documentation. 1.1 The first poverty assessment for Kosovo was done in 2001 on the basis of a Living Standard Measurement Survey of 2880 householdsconducted between September and December 2000. Although there was no existingcensus, effort was madeto create a representativesample of the populationof Kosovo. Up to date lists of households were created and a sample representative at areas of responsibility (AORs), ruralhrban, and AlbaniadSerbianethnicitywas drawn. 1.2 In June of 2002 Kosovo began to implement the Household Budget Survey (HBS). The HBS is implemented by the Statistical Office of Kosovo with technical assistance from Statistics Sweden, which in turn is financed by SIDA. To date four rounds of HBS have beencompleted(Table 1.1). SOK, together with Statistics Sweden, draw the sampleto be surveyedeach May. The first HBS survey began in June 2002 and 7 ran till May of the following year. The second survey (2003) followed the same cycle. But in 2005, SOK switched to calendar year (January to December of saymeyear) for the introductionof differences in the questionnairebut kept the timing of sampling the same at mid-year. Thus, currently, eachquestionnairespanstwo samples. 1.3 The Household Budget Survey provides a solid foundation for monitoring poverty in Kosovo. The H B S has become a core survey in KOSOVO~S efforts to build a long term monitoring and evaluation system. It has some of the basic tenets of a sustainable survey. It is fully funded by the government and implemented by the SOK staff (with technicalsupport from development partners). The H B S unit of SOK has also introduced innovations to the traditional H B S by including additional modules, most recent of which havebeenmigrationandremittances (2005) andtime use (2006). A. THEREARE PROBLEMS OFDATA COMPARABILITY 1.4 Examining changes in poverty and inequality over time in Kosovo poses several challenges. With a Living Measurements Standards Survey (LSMS)in 2000 and a series of H B S since 2002, it would seem tempting to conclude that tracking welfare changes in the first half of 2000 should be feasible. But there are practicalproblems. A major problem is that data are not comparable. There are three changes across surveys where efforts to compare data present difficultiesto trackingwelfare changes over time. Below we list each of these changes and discuss potential consequences for estimating changes in povertyand inequality. (a) Problem# 1:Diary versus Recall 1.5 The main change between HBS 2002 and subsequent HBS series is how households were asked to recall expenditures of goods and services bought. The first H B S asked households to record expenditures on a daily basis for two weeks. This appliedto food, own-produced consumption and most non-fooditems such as clothing, footwear, and education and health expenditures. A switch from a shorter to a longer recall period (diary to weekly) is likely to make households forget some details of consumption and therefore underreport consumption. The impact is likely to be severe for frequentlypurchaseditems such as food. (b) Problem#2: Survey Design-Redefinitionof ConsumptionItems 1.6 The secondchange which is likely to havean impact onthe comparabilityof data across H B S series is the level of disaggregation of the expenditure items. This took two forms. Inthe 2002 survey, householdsrecorded expenditure items on a blank sheet, but in subsequent years, the list was provided to the households. Between the first and second surveys, the lists did not exhibit substantial differences. It appears that households in the second survey were offeredthe same list that households interviewed in 2002 reported. However, by 2005, the level of disaggregation has increased and the list containedmore items. The more substantialchangewas in how consumption of own- produced items was reported. In this case, the items were aggregated into 12 categories (meat products, poultry, graincrops, and so on). Furthermore, inthe case of consumption of own-produced goods the recall period changed not from daily to weekly as in other items, but from daily to monthly. A shift from a smaller list to a longer list 8 (disaggregation) is likely to lead to higher reported consumption. (c) Problem #3: Survey Design LSMS versus HBS - 1.7 As the first post-conflict household survey, the LSMS estimates and profile o f poverty would be a good starting point to establish the baseline for the monitoring and evaluation system that is now anchored on HBS. However, except for the fact that the LSMS and the H B S are drawn from the same sample frame - that is, the households surveyed in HBS are selected from the same enumeration areas that were drawn for the LSMS - the two surveys differ in a number o f ways (Table A.1). First, distribution o f consumption may differ due to failure to account for seasonality in the LSMS. The latter was conducted for 3 months (September through December o f 2000), while H B S collects information from households (albeit different ones each month) throughout the year. If the three months when LSMS was fielded happen to be a period o f low (high) consumption, then the distribution o f consumption may be lower (higher) than the HBS distribution. Second, the recall period for consumption differs in the two surveys. In the LSMS, food and most frequently purchased non-food items had a recall o f a week or a month, while infrequently purchased goods and services had a 12 month recall. By contrast, the H B S first started with a two-week diary (daily recording) o f food and most non-food expenditures and then switched to a weekly recall. Finally, the LSMS provided households with a much narrower list o f expenditure items (46 food items) compared to HBS list that was over 100. Inpractical terms, a single change is hard to over come, but three makes the problem almost insurmountable. B. SAMPLE WEIGHTSINTRODUCE ADDITIONAL UNCERTAINTY 1.8 Sample weights introduce additional uncertainty. The Kosovo HBS uses the 1981census as the reference population. This is then updated every survey cycle through re-listing o f selected EAs, but it is not clear how the updated information is used in subsequent surveys. In addition to the outdated sampling frame, because o f resource constraints, field supervision o f surveys has been limited. Figure 1.1: Total Population in Millions and Household As a result, there is Size considerable uncertainty Sample Statistics surrounding the H B S *-I demographic statistics each year. shows the implied population count from the sample weights in the HBS. Within each wave o f the survey, it also presents the average household size o f sampled households by year and wave. The. estimated population appears to have 1 2 3 4 declined by about 25 percent between 2002 and 2005. Viewed from the perspective Source: World Bank staff calculations from the HBSdata. 9 o f this period, that is absence o f conflict and or unnatural mortality shocks, there is no clearjustification for this massive change in population estimate. 1.9 Strikingly, sometimes the surveys appear to come from completely different population groups. In particular, while average household size was near 7 in 2002, it drops to around 6 in 2005. Moreover, rural population shares change dramatically. For instance, the rural population decreases from 73 to 65 percent o f the total population. The 2001 LSMS reports rural population as 62.4 percent. The Agricultural Household Survey finds that rural population stayed at round 65 percent in 2004 and 2005. Based on experience in Albania (Carletto et al, 2004), we are expecting the incidence o f internal mobility to remain quite stable over time. One consequence o f this massive change in population estimate is to introduce huge volatility in the estimated count o f fraction o f people below the poverty line. Table 1.1: Population Size by Survey Wave and Year HBS estimates 2002-03 2003-04 2004-05 2005-06 Total population, inmillion 2.1 1.8 1.7 1.5 Number of households, inthousands 306.1 281.6 281 249.4 Householdsize 6.8 6.5 6.1 6.1 Rural as 'YOof total population 72.5 73.4 65.9 64.6 Reference population statistics 2001 - - 2002 2003 - - 2004 2005 Source LSMS LFS andAHS AHS AHS Total Population 1.97 1.9 Rural as % oftotal 62.4 Rural inmillion 1.23 1.3 1.3 1.3 Source: World Bank staff calculations from HBS data and LSMS: World Bank Kosovo Poverty Assessment (2001); Labor Force Survey(LFS) and Agricultural Household Survey (AHS) estimates are from the relevant SOK publications. C. LIKELYCONSEQUENCES:CONSUMPTION 1.10 Experiences around the world have documented the influence and magnitude o f the changes in recall period on consumption. In all cases, longer Figure 1.2: Average Monthly Household recall periods lead to less declared Consumption, in Nominal prices expenditures (Table 1.3). For instance, in India, households who were asked to report weekly food ' Average Monthly Expenditures expenditures had 15 to 20 percent higher per capita consumption than A those asked to report 30 day food expenditures, mainly because p households with shorter recall period reported higher per capita food cI expenditures (Tarozzi, 2002; Deaton, 2001). In another study, Deaton (2003) reports an experiment where reducing recall for food items period Source. World Bank staff calculations from the HBS data. from 30 to 7 days resulted in 30 10 percent higher consumption (1.1 percent per day). Amenuvegbe (1990) shows from Ghana household surveys that for 13 frequently purchased items, reportedexpenditures fell at an average of 2.9 percent for every day added. Lanjouw and Lanjouw (2001) showedthat variationsin food expenditure definitionsthat arise from a disaggregation of the list would lead to significant lower per capita consumption in countries such as Brazil, Ecuador, and El Salvador. For instance, fer capita monthly expenditures in El Salvador were 32 and 15 percent higher at the 10 and 90thpercentiles, respectively,for householdreceivingthe longlist. 1.11 Diagnostic work on Kosovo data indicates that expenditure data has been influenced by changes in recall period. The pattern is consistent with prior expectations as documented above in a number of other countries. It suggests, using Deaton (2003) results and notingthat food accounts for 50 percent of total consumption in Kosovo, that we shouldexpect at least 4 percent lower consumption in 2003 compared to 2002 from changes in recall periodalone (that is, 1.1 percentx 7 x 0.5).Inreality,we find that the mean of total consumption in 2002, which used the diary, was about 10 percent higher than the mean in 2003, where a weekly recall was used. It was15 percent higher than the mean in 2005. The mean of food consumption dropped by 13 percent between2002 and 2003, butby as muchas 21percentbetween2002 and 2005. 1.12 The effect of recall change may have been particularly severe for certain sub-components of consumption. As noted above, recording of own consumption underwent two substantial changes. One is the change in recall from daily to monthly. The other is that, in the second and subsequent surveys, households were given an aggregated list against which to record own consumption. More precisely, the list reported for own consumption changed from 85 in 2002 to 12 in subsequent surveys (Table 1.2). Both changes are likely to lead to underreportingof expenditures. Mean of own consumption fell by 4 percent between 2002 and 2003 and by 30 percent between 2002 and 2005. Giventhat small scale farmers -those with less than 3 hectares of land- report using 70 percent of their production for own consumption (SOK, 2005), the changes introduced in capturing this sub-component of consumption presents serious problems for a credible measure oftotal consumption,and ultimately,poverty inKosovo. 1.13 The possibility of survey design changes driving the changes in consumption (and therefore changes in welfare) cannot be ruled out. Food share fell from 61 to 54 percent between 2002 and 2005. In one view this could be an indication of households getting richer and substitutingaway from food to non-food. However, the evidence for this alternative hypothesis is not strong. First, the macroeconomic data shows a stable inflationregime (possibly even a deflation) and negligible output growth. Second, non- food expenditures remained stable across surveys in sharp contrast to food and its sub- components. Specifically, the share of sub-categories such as bread, meat or eggs and dairy out of total expenditures do not show evidence of substitution away from staples. Taken together, it appears that changes in recall period probably drive much of the observed changes, since as predicted these changes in recall periodare likely to have the biggest impact on frequently purchased items such as food. Simply put, since these changes in consumption (welfare) are observed in the context of several changes to survey design, it is difficult to argue credibly that observed changes are not due to changes in survey design. 11 1.14 Table 1.2: Summary of Survey Constraints and Their Effects on Poverty Estimates Survey and Possibleeffects References Evidence of effect in the Interaction and questionnaire HBS data final effect on design issues poverty Weak sampling Non- Demery and Population estimates: Interactswith all frame representative Grootaert (1994), 2002/03: 2.1 m other survey population. Howes and 2005/06: 1.5 m measurement Household size Lanjouw (1997) Rural proportion: errors. Leadsto and subgroups 2002103: 73% unquantifiable are not stable 2005/06: 65% (Table 1.1) biases. Change from Possibly an Currently no Own production drops by Poverty: open-ended to increase in controlled around30% from 02/03 to Underestimatedin close-ended reported experiments 05/06 05/06 or expenditure consumption (Volume I,Table A.1) overestimatedin questions estimates 02-03 Recall period Decrease in For survey, see Total food expenditure Poverty: change from reported Deatonand drops 21 % from 02/03 to Underestimatedin daily to weekly expenditureof Kozel(2005) 05/06 (Volume I,Table 05/06 or about 4%. A.1) overestimated in 02-03 Change in Decreasein Lanjouw and Own production drops by Interactswith number of reported Lanjouw (2001) around4% after the changes inrecall subcategories of expenditures and many others number of categories period and expenditures changes from over 85 to questiontype. reported 12 Cannot be singled out. Short recall Overstated Gibson (2005) Seasonality inpoverty Overstating period poverty estimates (Table B.4) poverty Source: World Bank staff calculationsfrom HBS data and relevant references. D. LIKELYCONSEQUENCES: POVERTYESTIMATES 1.15 A shift from diary to recall leads to underreporting of consumption, which in turn leads to higher estimated poverty rates. In 2002, the proportion o f people living below the poverty line was estimated at 37 percent. Using a consumption aggregate constructed in the same way and adjusted for inflation, the fraction o f the population below the poverty line increases to 44 percent in 2003, fell to 35 and increased back to 45 percent in 2005. 1.16 Viewed differently, the disaggregation o f consumption items is akin to introducing measurement error into a variable (Table 1.2). If the measurement error i s random, there will be no effect on the estimates o f the mean or the population total if the sample i s large enough. However, such errors will systematically bias poverty estimates. Figure 1.3 shows a situation where an accurate welfare indicator i s compared with an error-ridden indicator. The poverty rate is the area under the welfare function up to the poverty line and it will be affected both by imperfectly measured welfare indicator, or incorrectly specified poverty line. 1.17 12 Figure1.3: PovertyRateEstimatesand the Effectof Changesinthe Questionnaire A. PovertyRatesOver SurveyPeriods, B. TheEffectof RandomMeasurementError Absolute, Extreme on PovertyEstimates Source: World Bank staff calculations from HBS data. Source: Gibson (2005). 1.18 Samplingweights increase the volatility of the estimated poverty. Table 1.3 compares the estimated poverty rates with and without weights. A comparison of the weighted and un-weightedcolumns shows why using weights as currently constructed introduces volatility. The magnitude of changes is further overstated with the weighted statistics. For instance, for urban areas, the weighted poverty rates seem to drop by 5 percentage points whereas the un-weightedby only three. For rural, the value of the supposed increase in poverty is much smaller when the sampling weights are not included. These findings suggest the need for a consistent procedure for calculating samplingweights. Table 1.3: PovertyHeadcountby Locationand Ethnicareas, usingPA05 methodology Weighted Unweighted 2002-03 2003-04 2004-05 2005-06 2002-03 2003-04 2004-05 2005-06 Total 37.7 43.7 34.8 45 45.4 44.5 34.4 44.3 Rural 34.4 44.2 37.2 49.2 42.1 46.2 37.5 49 Urban 46.6 42.1 30.3 37.4 48.1 42.7 31.4 39.7 Albanianarea 37.8 43.8 34.9 43 45.3 45.1 34.5 41.7 Serbianarea 33.5 40.8 33.3 80.4 44.1 38.4 33.8 70.3 Source: World Bank staff calculations from HBS. Notes: Methodology as in the 2005 Poverty Assessment. Weighted refers to individual-level weights, unweighted to householdsize weights. 1.19 These uncertainties persist across several estimates. In additionto nationallevel estimates by wave, poverty rates were estimated for rural and urban residents and Albanian and Serb ethnic groups. For instance, estimates of poverty by ethnicity, whether definedas area occupied mainly by such an ethnic group or ethnicity of headof household, are highly volatile. For instance, the povertyrate for Serbs ranges from 35 to 80 percent. They are especially sensitiveto inclusionof own consumption. For instance, in where we present the poverty rates under different consumption aggregation with the same poverty line, the coefficient of variation (the standard deviation over the mean) of the poverty rate increasedwith the inclusionof own productionfor weighted figures. In 13 all cases, these problems of large changes betweenweighted and unweighted, and within a short time period, are observed. 1.20 The data from 2004-05 (wave 111) seems to be particularly problematic. This survey was done in the same way as waves I1and IV (that is, 2003-04 and 2005-06) so that in theory it should be comparable to these surveys. However, we find that it is particularlysensitive to the inclusionofconsumptionof non-food. The estimatedwelfare swings with and without inclusion of non-foodare (unrealistically) large. This leads to the conclusion that estimated poverty counts are not comparable, especially between 2002 and 2003. Inthe next chapter, we try to resolve this issue in a number of ways and providepreliminaryestimates ofpovertytrends inKosovo. Table 1.4: Poverty Headcount by Household Head Ethnicity Weighted Unweighted 2002-03 2003-04 2002-03 2003-04 2002-03 2003-04 2002-03 2003-04 Albanian 37.4 43.8 32.1 42.5 45.2 45 32.9 41.1 Serbian 30.1 36 34.3 81.7 39.1 35.7 33.8 70.2 Other 57.6 53 67.2 51.7 58.2 59.5 57.6 56 Source: World Bank staffcalculationsfrom HBS. Methodologyas inthe 2005 PovertyAssessment. Weighted refers to individual-level weights, unweightedto householdsize weights. 14 CHAPTER2: POVERTY-ALTERNATIVE ESTIMATES 2.1 The dual problem of (i)possible survey bias inthe data, and (ii)numerous changes in questionnairedesign, make HBS survey estimates merelysuggestive of a trendand shouldbe used only as a guide by policy makers. Numerous changes in survey design do not leadto conclusive comparisons on the levels and trends in poverty between 2002 and 2005. We have shownthat a shift from diary to a weekly or longer recall period,from 2002 to 2003 and thereafter, respectively, is likely to leadto underreportinginconsumption and therefore over- estimationof poverty rates. We have also discussed that aggregation of own consumption items from 85 to 12, in 2002 comparedto 2003 and thereafter, adds to the underreportingof consumption (and by consequence over-estimation of poverty) problems in second and subsequent waves. Finally, the sampling methodology, which indicates a larger population and higher household size in 2002 compared to 2003 and thereafter, is likely to reduce per capita consumption and, for a given poverty line, under-estimatepoverty in 2002 relativeto 2003 and thereafter. While we know the possible direction of impact of these changes in design on consumption and poverty, it is not possible to know with precisionthe magnitude of these changes on consumption or poverty. That is why, the searchfor alternativemethods to establishcomparabilitybecomesnecessary. 2.2 We employ several estimation techniques to correct for some of these problems. In order, we present a brief descriptionof the steps taken to address the (a) sampling issues and(b) non-comparablewelfare measures. (A) Sampling issues: As the HBS data is based on an outdated sample and the survey supervision is very limited, the data suffer from a possible bias. To rectify a part of this problem, we use a post-stratificationprocedure. This method calibrates the weights to make demographic estimates from HBS comparable to external sources. 2.3 Even if sampling issues are addressed, the problem of non-comparability of consumption estimates still persists. Therefore, we apply the following steps to rectify this secondproblem: (B) Comparability of welfare measures: We use two main methods to provide comparableconsumption aggregates. Compare only 2003 and 2005: Since the biggest and the most problematicchanges took place between2002 and2003, one strategy is to ignorethe 2002 survey and start the analysis of poverty from 2003. As a reminder, the 2003 through 2005 data have the same recallperiod. The level of item didaggregationcan also be considered the same, since only minor changes were introduced. For instance, food items declined from 114 to 107 between 2003 and 2004, and similar changes were introduced in non-food items. But overall, the number of the changes in consumption items and their contribution to aggregate consumption were negligible. Our justification for 15 excluding2004 survey is that welfare changes are very sensitive to inclusionof non- food consumption. Therefore, we use three methodologies to compare poverty between2003 and 2005: 0 First, we use the same constructionof consumption aggregate and poverty line as was used for the previous two Poverty Assessments. We refer to this as PA05 methodology(short for PovertyAssessment 2005). Then, we directly compare the povertyrates. 0 Second, a method developed by Lanjouw and Lanjouw (2001) is used to construct a Comparable Consumption Aggregate that includes only consistently recordedexpenditureitems and least volatile items. 0 Third, we constructan Abbreviated ConsumptionBundleconsistingonly of products for which price informationwas collectedby the price unit of the SOK in order to re-calculatethe poverty linefor 2003-04 data 0 Compare all the years: The final optionis to compare all the years. However, as argued above, this cannot be done without additionaladjustment. There are two candidatemethodsfor adjusting povertyratesto arriveat comparability. 0 The first method, calledinverseprobability weighting, aims to matchthe distribution of consumption or any welfare measure between the two surveys. It reweighs the poverty count in 2002 using as weights the probability of an observation belonging to a comparison survey, say year 2003 (Tarozzi, 2005; DiNardo et. al. 1996). Similarly one can compare 2002 to 2004 and2002 to 2005. 0 The second method, which we shall refer to as econometric projection methodology, is to estimate a consumption modelusingthe 2002 data, and then use the estimated parameters from the 2002 model to forecast or predictthe consumptionfor subsequentyears. The final step is to addto the forecast consumption an estimate of unobserved part of consumption (the error term) in order to recover full consumption. The results from this methodology are notyet complete and are notreportedhere. 2.4 All methods suggest that the povertyrate in Kosovo remained in the mid 40s percent from 2002 to 2005. The results of the different estimation methodologies are presented in (Table 2.1). Although the trends are not consistently pointing to the same direction, the pattern that emerges is one of stagnating poverty. The PA05 and abbreviated consumption methods imply a stagnant poverty rate: a change from 44 to 45 percent. The Comparable Consumption Aggregate, suggest a slight decrease from 49 percent in 2003-04 to 46 percent in 2005-06. The Inverse Probability Weighting methodology also confirms that poverty remained very similar from 2002-03 to 2005-06, with only a small increase of about 3 percentage points from 2003 to 2005. These conclusions do not change substantively if survey waves are defined differently. Specificallywhen usingcalendar year 2005 as the last wave, the results show only a small decline in poverty. See Annex B and particularlyTable B.5. 16 Table 2.1: Overview of the Results of Methodologies for Comparable Poverty Estimates ConsumDtion Aggregate (CA) Povertv Line definition definition Poverty Rates CA with PA 05 methodology Poverty line 2002 adjusted with CPI weighted unweighted 2002-03 37.6 45.7 2003-04 43.6 44.5 2004-05 34.8 34.4 2005-06 45 44.3 Comparable CA (Lanjouwand Robust Poverty Line Lanjouw,2001) weighted unweighted 2002-03 2003-04 48.5 48.7 2004-05 41.3 41.3 2005-06 45.6 44.7 CA IUsing Inverse Probability 2002 Poverty Line Weighting (Tarozzi, 2005) weighted unweighted 2002-03 42.7 2003-04 2004-05 2005-06 45.6 Source: World Bank staff estimates from HBS data. 2.5 The results presented in the main part of the report (Volume I) for 2003/04 and are 2005/06 only, re-weightedto match non-HBS based rural and urbanpopulationestimates. In this volume, we present the results from the methodologies discussed above in order to see whether the mainresultofVolume I, of unchangingpovertytrend, is confirmed. Inshort, that inthis volume we undertakea sensitivityanalysis. Table 2.2: Summary of Poverty Estimates from the Methodologies Used Methodology Base year Final year Change Poverty rates A. Samplingissues 2003104 2005106 1. Post-stratification 43.5 45.1 About the same B. Comparability of Welfare Measures Compareonly 2003104 and2005106 2003104 2005106 2. PA05 43.6 45 About the same 3. ComparableConsumptionAggregate 48.5 45.6 Slight decrease 4. AbbreviatedConsumptionBundle 36.3 36.2 The same Compare all surveys2002-2005 2002103 2005106 5. InverseProbabilityWeighting 42.7 45.6 Slight increase Source: World Bank staff estimates from HBS data. A. POST-STRATIFICATION 2.6 Because of an outdated sampling frame and resource-constrained limited survey supervision, the H B S sample is likely to be affected by non-negligible sampling and non- sampling errors. As discussed in the survey sampleseach year appear to come from different populations. This is reflected also in the distribution of the consumption aggregate. As 17 Figure 2.1 shows, the cumulative distributions of consumption from year to year even indicate a stochastic dominance of 2005106over wave 2002103 and 2004105. This figure also shows the similarity of 2003104 and 2005106 and wave 2002103 and 2004105. We suspect that these patterns may be driven by both changes in the questionnaire and the sampling procedure. Figure 2.1: Cumulative and Density Distribution of Consumption for the Bottom 50 percentile of the Population C ionof n Source: World Bank staff estimates from HBS data. 2.7 The sampling process and survey administration is poorly documented. The quality of the list of EAs is poor:the distinctionbetweenurban and rural is purely administrative;the classificationby ethnicity does not follow strict rules, and the descriptionof the geographical boundaries of the EAs is outdated (Andersson, 2002a). In addition, due to Table 2.3: Poverty Rates with Current Weights lackproper supervision misclassified Wave Population Averak and Rewei hted Of Extreme Absolute EAs were skipped (Andersson, 2002c), estimates household poverty poverty relisting of large EAs may be (million) size rate ('YO) rate ('YO) incomplete and field control of Current SamplingWeights enumerators is lacking. Some areas that 2002-03 2.05 6.8 15.0 37.8 were heavily populated in 1981 are 2003-04 1.82 6.5 13.3 43.7 currently not and vise versa This 2004-05 1.71 6.1 10.7 35.5 introduces large sampling errors and 2005-06 1.52 6.1 16.5 45.6 possibly bias to the HI3S estimates. Reweighted There are also issues of undercoverage. 2002-03 1.9 8.5 15.3 38.7 2003-04 1.9 8.0 13.6 43.5 2004-05 1.9 7.4 10.6 34.8 2.8 The reweighting methodology 2005-06 1.9 7.7 16.7 45.1 adjusts the sampling weights attached to each surveyed householdso that the urban and rural populationmatch non-HBS based data Generally survey data and its sampling weights are re-calibrated and post-stratification weights are used to match the distribution to some external data (Lohr, 1999). The adjustment methodology is simple and it uses a scaling factor so that the weighted total population size in all surveys matches that of external sources. Then it also matches the distribution of rural and urban households as compared to that of other surveys (Table 1.1). The resultingweighted populationtotal and household size is muchmore comparable (second half of (Table 2.3). We also match household size distribution in each stratum and obtain very similar results. 18 2.9 The re-weighted poverty rate confirms the time trend of unchanging poverty over time, while the volatility ofthe estimateshas decreased. The poverty rate, when re-weighted, i s again around 45 percent for 2003-04 and 2005-06. At the same time, its decrease in 2002 and 2004 is smaller than when calculated without post-stratification. This procedure, however, seems insufficient in equating the samples. As next steps, the analysis will adjust for other aggregates on which official data is available, as for instance pensioners and students. Box 2.1: Bosnia and Herzegovina HBS: Example of Sampling without a Census Bosnia and Herzegovina's HBS sampling faced similar constraints to those of Kosovo. First, there were no populationregisters or housing registers to be used as sampling frames. Second, there was possibly considerable internal migrationand rapid change amongst the housingstock. Third, the statistical office staff had limited resources and little experience of general population sampling methods (Lynn, 2004). The Bosnia and Herzegovina HBS sampling process follows the steps identified in Table 2.4. The procedure is similar to what currently SOK employs except for several noteworthy differences: census EAs are well delineated and stratified; relisting and questionnaire administrationis better supervised; use ofequal probabilitiesbothat the stage of selectingPSUs and at the stage of selecting households within PSUs. Table 2.4: Sampling procedure for the Bosnia and Herzegovina's Household Budget Survey Stage of Steps Time sampling Implemented only once Pre-sampling Revisedthe census EAs to ensure comprehensivenessand 5 months appropriatemaps Fieldtest A systematic randomsample of 50 EAs to find percent of 1 month unoccupieddwellings. Implementrelisting procedure and follow up visit. Implemented before the survey each year 1st stage Systematic equal-probabilitystratifiedsamplingof 3.65% of EAs. Relisting Semi-intrusiveapproach(observationwhere possible, contact 3 weeks elsewhere). About 1 day visit per EA. 2nd stage Systematic selection of Households from the relist (about 25% of all relisted). Systematic division ofthe sampledhouseholdsinto 12 monthly samples Source: Lynn (2004). B. COMPARE ONLY 2003 AND 2005 PA05 Methodology 2.10 The poverty rate is around 45 percent inboth 2003-04 and 2005-06 with a substantial decline in 2004-05 that is as yet unexplained. We use three methods to compare the poverty rates between 2003 and 2005. The first method uses the same poverty line used for the poverty assessment of 2005 (PA05), adjusted for inflation to estimate the poverty rates. A comparison o f all three years shows that poverty levels remained stagnant between the start and end o f the period. The poverty rate was at 44 percent in 2003 and 45 in 2005. But in 2004105 there i s a large drop in poverty, to 35 percent. While the pattern of change is consistent with the macroeconomic developments - there was a 2.6 percentage point 19 turnaround in GDP growth between 2003-04 and 2004-05 --such a decrease over a short period of time implies unusually high growthelasticityof povertyreduction'. Table 2.5: Poverty Rates with the PA05 and Comparable CA methodologies PA05 Comparable 2.11 Although the last 3 HE3S surveys ' methodology CA appear very similar and seem to be prime Povertyline 2002 PL Robust PL candidates for comparable poverty Povertvrates estimates, changes in the aggregation of 2003-04 43.7 48.5 food items could affect the poverty figures. 2004-05 34.8 41.3 The 2003-2005 HBS surveys usedthe same 2005-06 45 45.6 recall period. Generally, there is a Source: World Bank staffestimatesfrom HBS data. presumption that the groups surveyed are similar: the samples were drawn from three adjacent time periods, between which there had been no expectation of a marked change in poverty. However, they used different levels of aggregation: for instance, there are 107 food items in 2005-06 and 114 in 2003-04 and 2004- 05 surveys. Several additional non-food consumption items were added. Possibly, the changes in survey design produced a (misleading)appearanceof a drop and then an increase in poverty. The 2004-05 is particularlyproblematicand as hasbeenmentionedvery sensitive to inclusionofconsumptionof non-food. C. COMPARABLE CONSUMPTION AGGREGATEMETHODOLOGY 2.12 The second method, which adjusts the poverty line to account for survey-design induced volatility of consumption sub-components shows a slight decline in poverty. We noted that consumption and welfare estimates for 2004/05 survey were noticeably more sensitive to inclusion of consumption of non-food. To address this concern, we use a methodology (Comparable Consumption Aggregate) which constructs the poverty line each year. First, we construct a food poverty line for a reference population using only comparable food consumption items. Then we construct an absolute poverty line each year, non-parametrically(see Box 2.1). 2.13 The differences between the robust and the poverty line from the 2005 Poverty Assessment (Table 2.6) is not only the result of inflationover the period,but also reflects the fact that the 2005 survey embodies a more comprehensive consumption definitionthan 2003 and 2004 surveysas well as the issues arisingfrom biasedsampling and measurementerror in the second half of 2004. On the basis of these robust poverty lines, the incidence of poverty in Kosovo decreased slightly from 48 percent in 2003-04 to 46 percent in 2005-06. This contrasts with the observation that poverty increasedslightly from 2003 to 2005 when only inflation is adjusted for. In addition, the magnitudeof the drop between 2003-04 and 2004- 1 Most likely, the reported higher expenditure by households i s due to survey administration and sampling issues. As shown in the previous sections, the survey methodology could be introducing an unquantifiable bias. Measurement error is also a big concern for the Kosovo HBS as described earlier. Because of limited resources and capacity, survey administration is not at par with international standards: enumerator supervision is compromised while the incentives for respondents changed. This unknown measurementerror poses a special challenge when the focus is on poverty and other distributional statistics, rather than on means and totals. While random measurement error should not affect estimates of the mean or the populationtotal if the sample is large enough, such errors will systematically bias poverty estimates (Gibson, 2005). For poverty rates and other variance-based statistics, the effect of random errors accumulates so errors in measuring household level welfare will be reflectedin inaccurate estimatesof aggregatepovertyrates. 20 05 is much smaller than when consumption o f own-production is included. The trend now shows that poverty declines from 48 percet to 41 percent between 2003 and 2004. Table 2.6: Robust PovertyLinesBasedon Consistent Food Items. Food Poverty Line ExcludedOwn Production 2003-04 2004-05 2005-06 RobustFoodpoverty line 26.2 27.11 22.75 Robustfinal poverty line 41.84 44.2 1 40.39 CPI adjustedPA05 food poverty line 28.35 28.35 28.34 CPI adjustedPA05 poverty line 43 43.01 43 Source: World Bank staff estimates from HBS data. In Euros per adult equivalent, monthly, in June 2002 prices. AbbreviatedConsumption BundleMethodology 2.14 This fourth methodology re-calculates the poverty line for 2003-04 data (second wave) usingnon-HBS price information of 40 items. The calculation of poverty line is based on the household total consumption of certain reference population. Thus, the poverty line calculated for 2002-03 data in the 2005 Poverty Assessment is based on the consumption recorded in 2002-03. As we pointed earlier, consumption in 2002-03 was recorded using a diary method and it is different from later years. Unfortunately, for 2003-04 survey no price information was collected that can allow us to replicate the poverty line for that data Using non-HBS price information we are able to calculate the cost of an abbreviated consumption bundleof 40 items. Table 2.7: PovertyRates usingthe Abbreviated ConsumptionBundleMethodology Survey Adjusted Adult Food line Complete Extreme Complete wave Equivalent Poverty line Poverty Poverty Consumption Rate Rate 2003-04 52.26 22.39 39.01 5.85 36.28 2004-05 59.73 21.84 38.05 4.84 27.23 2005-06 52.12 21.83 38.03 8.89 36.19 units Euro/month Euro/month Euro/month % YO Source: World Bank staffcalculationsfrom HBS data. Wave 2 Poverty line is recalculatedusing40 major food item. The povertylinesfor waves 1,3,4 are deflatedfrom wave 2 povertylines usingCPI. 2.15 Based on these new poverty lines, poverty rates in Kosovo remained stagnant from 2003 to 2005, thus confirming results from other methods. The poverty line is lower than the one calculated for 2002 since it is abbreviated. The poverty line calculated usingH B S price information for the 2002-03 data was 43 Euros per month, while this one is 22 Euros per month. Thus the poverty rate appears to be lower. The lower poverty rates are not driven by any real changes in the welfare but simply by this estimation technique. It i s the poverty trend that is informative. The resulting poverty trend confirms findings from other estimations that poverty rates remained stagnant. 21 Box 2.2: Analysis of Changes Analysis of changes in poverty presentedhere is based on consumption data from the 2003-2004, 2004-05 and 2005-06 Kosovo Household Budget Surveys. The consumptior modules differ over the survey waves: the 2005 HBS included more items than the 2003 an( 2004 surveys. Because the consumption modules differed it was necessary to put together i comparable consumption aggregate (CCA) with each survey. The CCA is a single consumptior value in each survey, constructed such that the sets of components inthe aggregate in the 2003 2004 surveys and the 2005 survey are parallel. Because the CCAs were assembledsolely for thc purpose of maximizing comparability across the two years, the CCA is not identical to the ful consumption aggregate used inthe first part ofthe report, Volume I. Following the methodology developed by Lanjouw and Lanjouw (2001), we define ar abbreviated food poverty line based only on the categories included in the CCA. (Given thc differences between the CCA and the full consumption aggregate, it would not be sensible tc apply the poverty lines based on the full consumption aggregate to the CCA (see Table B.5 Poverty Rates Using Alternative Consumption and Poverty Line Methodologies). The fooc poverty line, z, is defined as the average expenditure on these comparable items by thc populationinthe 30 to 50 percentiles(26.2 Euros for 2003-04). The robust final poverty line, Z derived from this abbreviated food poverty line is 41.8 Euros for 2003-04 surveys and 44.2 an( 40.4 Euros per month for 2004-05 and 2005-06 surveys respectively. Each line is calculate( non-parametricallyby taking average total consumption among sample households with fooc expenditure within 1 percent of z, within 2 percent of z, in increasing bands to within 5 percen of z. The final poverty line, Z, is thenthe average o fthese values. The values are listed in Tablc 2.6. A major assumptionbehindthis methodology is that expenditures on the goods includec inthe CCA have an Engelcurve relation to more comprehensivemeasuresof expenditure. Engel's law postulatesthat the higher the total expenditure, the lower the share of food expenditures. A major assumptionbehindthis methodology is that expenditures on the goods includec in the CCA have an Engel curve relation to more comprehensive measures of expenditure Engel's law postulates that the higher the total expenditure, the lower the share of fooc expenditures. This assumption appears to be met with this data. Other assumptions that needtc be satisfied for this methodology to be robust are stable expenditure patterns and no mis measurement inthe data. The other requirement for the comparisons to be robust is that only thi head count measure of poverty is used. The problemwith higher order poverty measures is tha the relativedistance betweenthe consumption levelof the poor and the poverty line may increasl as the components in the consumptionaggregatebecomemore comprehensive. It should be emphasizedthat the fact that the two surveyswere not identicalmeansthat the CCAs at best are only approximately comparable. As a result, the use ofthe CCAs introduces a levelofunquantifiableerror beyondthe usual sample error. Thus, the apparent changes over time should be interpretedwith caution D. COMPARE ALL THE YEARS 2.16 The procedure employed in this section involve estimating an econometric relationship between welfare and household characteristics with the 2002-03 data, using a set of characteristics common to all surveys. The estimated relationship is then used to update the distribution of the explanatory variables in the later surveys with information on the conditional probability (the estimated relationship) from the 2002 survey (Inverse Probability Weighting (IPW)*. 2The procedure used here is very similar to that of Stifel and Christiansen (2006), drawingheavily on the work of Elbers, Lanjouw, and Lanjouw(2003). 22 InverseProbabilityWeighting 2.17 The IPW consist of two estimation steps that corrects for the difference in the distribution of consumption between two surveys. In the first stage, data for 2002 and a comparison year are combined and a logit or probit model estimated where the dependent variable is 1 if an observation belong to year 2003 and 0 otherwise and the independent variables are a set of variables that have not changedfrom survey to survey. This enables us to obtain the predicted probabilitythat an observation is part of 2003 (propensityscore). In the second stage, the estimated propensity score is used to reweigh the poverty counts in 2002. The reweighted poverty estimate provides a comparable poverty count for 2003. By doingthe same thing for 2004 and 2005, we obtain a series of poverty counts all comparable to 2002 data. 2.18 The estimation results from using IPW methodology also suggest that poverty headcount increased. Our preliminary results using data from 2005-06 and 2002-03 only suggest povertyoutcomes probablyremainedthe same. The results indicatea slight increase, but the magnitudeofthe increasedepends onthe householdcharacteristics specified in step 1. The increase is from 42 to 45 percent between 2002 and 2005 if a large set of household characteristics are used (Table 1.2). If we employ a more limited set of household characteristics,thenthe impliedincreaseof povertyis higher. E. COMPARISON OFPOVERTYFIGURES FROMTHE LSMSAND HBS 2.19 LSMS and HBS data are not directly comparable because of differences in item definitions, disaggregation and recall periods. The 2001 Poverty Assessment (PA) reported the poverty rate at 50 percent usingthe LSMS 2000 data, while in 2005 the PA reported this rate at 37 percent using the HBS 2002 data. While both reports estimated household expenditures and a relevant poverty line, their results are not comparable becausethe LSMS and HBS surveys differ significantly in their representativeness and survey design. The representativeness of the surveys is difficult to gauge because of lack of a recent population census. In addition, the sampling frame of the LSMS was revised for the HBS. There are substantial disparities in the populationestimates for urban and rural areas as well as for the ethnic groups (Poverty Assessment, 2005). 2.20 The questionnaire designs preclude the constructionof comparable poverty indexes. The main differences are in the expenditure item definitions and in the recall period. However, in theory, the inverse probability weighting methods should apply here as well. There are also a few methodologicaldifferences in the constructionof the poverty indicesin these two poverty assessments. The main methodologicaldifferences are the exclusion of durables and the inclusionof health expenses for the consumption aggregate usingthe H B S (Table 1.2). By comparison, the PA using LSMS included durables but excluded health expenditures.Another differencein the estimation of poverty numbers in the two PASstems from the need to account for survey design when calculating point estimates. The H B S survey is stratified at the urbadrural and ethnic group level, but the 2005 poverty estimates did not adjust for this pattern of stratification. It is therefore difficult compare how large a difference the two estimates are from each other. The per-adult equivalent consumption aggregate andthe poverty lines are otherwise constructed inthe same way. 23 CHAPTER3: CONCLUSIONSAND RECOMMENDATIONS 3.1 The HBS demographic estimates suggest that there is an unquantifiable bias due to the outdated sampling frame and limited survey administration. The survey data implies incorrectlya populationreductionfrom about 2 million to 1.6million as well as very volatile urbanandrural dimensions. These estimates are suggestiveof a sampling biasthat cannot be correctedfor without proper sampling frame. The currentlysampling frame for the HBS data (and other SOK surveys such as LFS) dates back to 1981. A sample drawn from the 1981 frame will approximate the population in that frame. The differences between the 1981 populationdistributionand the current are introducingvery large sampling error and possibly bias. Inaddition,the survey administrationhasvery limitedfield supervision. 3.2 Changesin the survey questionnaireadditionallymake povertynon-comparable between years. The H B S questionnaire and how it asks respondents to report their consumption changes dramaticallyfrom 2002/03 to 2004/05 and then additional changes are introducedeach year. As the literatureon the subject prove, changes in the recall periodand the disaggregationofthe categories producedifferent consumption estimates even when there are no consumptiondifferences in reality. 3.3 We employ six methods to address the issues non-comparabilityand sampling. For sampling, we employ post-calibrationto match HBS estimates to other estimates of the population. We also exclude volatile waves, as for instance the consumption patterns from 2004/05 seem to not be comparable to those of 2003/04 or 2004/05. To address non- comparability of how the expenditure data was collected, we use, first, surveys with comparablequestionnaires, and second, all surveysbut correct for differences in consumption patternsby usingeconometric projectionand inverse probabilityweighting. 3.4 Based on this extensive sensitivity analysis, we do not find firm evidence of improvement of householdwelfare in Kosovointhe last three years. The resultthat is robust to specificationand different methodologiesis that there is no significant increaseor decrease inthe povertyrate in Kosovofor the periodfrom 2003/04to 2005/06. A. RECOMMENDATIONS 3.5 Consistencyinthe survey questionnaire should be the goal of next surveys. Even small changes in the way how questions about expenditures are asked can cause differences in reported consumption even if there is no real change in the consumption pattern of the household. Changes in the survey questionnaire should not be introduced without a randomizedexperiment beforehand. A randomizedexperimedmini-surveywill allow to test the effect ofchangingthe question on the reportedexpenditure. 3.6 A census is urgently needed to create a basis for unbiased sampling. It is unfortunatethat four years of data are not reliableenoughto providepolicy guidancebecause of a lack of census. A census will completelyresolve the sampling issues for future surveys, andpossiblycan be usedretroactivelyto adjust earlier samplingweights. 25 3.7 Better survey administrationand documentation of all steps of the process and the data are necessary. Currently, the survey administrators have hardly any supervision. Additional supervision will affect the HBS budget only marginallybut will have highreturns in improveddata quality. Second, documentationof all procedures is very limited. This step is at least costly, but will improve data quality and usefulness. 26 ANNEX A: TABLESAND FIGURES Table A.l: Comparison of Previous Methodologies WB Poverty Assessments 2001~ 20054 Poverty Absolute Extreme Absolute Extreme Rates Total 50.3 11.9 Total 37.0 15.2 Rural 52.0 11.6 Rural 34.1 14.8 Urban 47.5 12.5 Urban Not avail. Pristina 36.4 7.7 Other 47.1 19.1 Data LSMS 2000 HBS2002-2003 Timing: Sept -Dec 2000. Timing: 612002-512003. Sampling frame: Sampling frame: 2, 880 households. 2400 households Rural: basedon the Housing Damage Similar to LSMS with primary sampling Assessment Survey (1999). units revised. Urban OSCE voters' registration. Representativeness:Urbadrural; ethnic Representativeness:Urbadrural. AORs, group. ethnic groups. Consumption Food: Aggregate 1. Purchasedfood inthe last 30 days Food: in39 categories, both quantity and 1. Expenditures on food. value. 2. Own production 2. Stored food inthe last month and 3. Foodout last year, 7 categories. 3. Own production, gifts. 4. Food out. Housing Expenditure and Rent: Housing Expenditure and Rent: excluded. excluded. Non-food: Non-food: personal items, hh services. Personalitems, hhservices Semi-durables Durable goods: Durable goods: rental value. excluded; Education: Education: included, 1-year recall. included, diary method Health: Health: excluded included Price indexes: using unit values (ratio of Price index: values over quantities) after excluding obs > CPI by monthand urbadrural dimension. Information here i s from the Poverty Assessment, 2001 x-1Appendix G of the Kosovo LSMS 2000 Basic Information Document from http:l/www.worldbank.org/lsms 4Kosovo Poverty Assessment (2005) and Tsirunyan, Sasun. 2004. "Poverty and Inequality in Kosovo", backgroundpaper for the Poverty Assessment. 27 2 st.d. Paasche price index. Equivalent EA, = (A, + BC,)'where B= 0.75. EA, = (A, + BC,)'where e= 0.75. Adults Equivalent Adults = (Adults + -75 Equivalent Adults = (Adults + .75 Children),". Children < 15 years old. Children),". Children < 15 years old. Per Adult- Equivalent pEC, = TCI TCI (A, +0 COY X(A0 +6 cole PEC, = Consumption (A, +.ec,)' A, +c, ( A ~ c,)' +e A, +c, where the pivotal householdhas 4 adults where the pivotal householdhas 4 adults and 2 children (Ao=4, C0=2). and 2 children(Ao=4, C0=2). Poverty Food poverty line: Same methodology as for 2001. Food Lines Basedon 2, 100 calories per adult. Caloric basket of 2100 calories is estimatedwith the structure of the 30" to 50" population price information from the HBS. Caloric percentiles. structure of the 30* to 50" population Food line: percentilesfrom the HBS. DM 1A529 per adult per day. Food line: Poverty line: Euro 0.93/day/adult. DM3.498 per adult per day. Usingthe Poverty line: share of non-food items for hh with food Euro 1.41per adult per day. consumptionclose to the FoodLine. Food Same methodology as 2001. Food share= share= 53.97%. 65.9%. Currency Lack of PPP adjustment indexes. PPP not available. conversion Currency conversions use the rates Not indicated. correspondingto the month of the survey. Unofficial exchange rate of 30 to 33 Dinars per DEM. Other Not accounting for stratification. Source: World Bank KosovoPovertyAssessment (2001) and PovertyAssessment (2005). 28 Table A.2: Survey Comparison KOSOVO HBS Wave 1 Wave I1 Wave I11 Wave 1V HBS-2002-2003 HBS-2003-2004 HBS-2004 HBS-2005 Period usedfor analysis 612002-12003 612003-612004 612004-512005 612005-512006 Number of 2400 households 2400 2400 2400 observations ((960 rural, 1440 urban) households households households Survey questionnaire design and its changes Timing of questionnaire 612002 612003 112005 112006 introduction Food consumption expenditure Recall period daily weekly weekly weekly Method diary recall recall recall Questiontype open-ended close-ended close- close ended ended Categories 165 103 103 107 Consumption of own production Quantities no no no In-kind food receivedas Yes Yes Yes gifts, donation Categories 85 12 12 12 Non-food expenditures Education daily diary weekly recall weekly weekly recall recall Categories 14 13 13 13 Health daily diary weekly recall weekly weekly recall recall Categories 6 6 6 11 Other non-food Clothing daily diary weekly recall weekly weekly recall recall Categories 31 10 I O I O Householdtextiles yearly recall weekly recall weekly weekly recall recall Categories 6 6 6 6 Transport daily diary weekly recall weekly weekly recall recall Categories 11 5 5 15 Durables Purchases Yes Yes Yes Ownership quantity of item no no Value no no no When bought no no no Housing consumption Rent no Yes Yes Categories 2 6 6 Estimatedrent ifowned Yes Yes Yes Utilities daily diary weekly recall weekly recall weekly recall Categories - 16 24 24 26 Source: Relevant HBSquestionnaires and datasets. 29 Table A.3: Percent Changt in Main Aggregates from Survey Survey Comparison Changefrom ..... IItoIII IIItoIv IItozv Base is 03/04 04/06 03/04 TotalConsumptionofHH 9% -13% -5% TotalExpendituresof HH 14% -13% -1% Consumptionof ownproduced -18% -12% -28% or fetchedfood Food expenditures (incl. 2% -12% -9% alcohol andtobacco) Non-Foodexpenditures 32% -15% 12% Source: World Bank staff estimates from HBS data. Table A.4: Alternative Consumption Aggregate Definitions and Poverty Rates Poverty Rates Coefficient ConsumDtionAggrePate Weighted of SDecification Variation 2002103 2003104 2004105 2005106 2002-05 Basic food excl ownprod. plusbasic 77.8 81.7 73 79.7 5yo non-foodspendingexcl. utilities Above plus ownproduction 61.3 68.8 58.8 69.6 8Yo Above plus in-kind, food out, alcohol 55.6 63.5 51.4 62.6 10% andtobacco Above plus semi-durablesandutilities 46.4 49.6 38.9 50.1 11% Above plus education 46.4 49.6 38.9 50.1 11% Above plus medical 44.6 47.6 36.8 48.4 12% Source: World Bank staff estimates from HBS data. 30 Table AS: Consistently Asked Questions over the 4 Surveys Variable Wave I 2003-04/III Wave IV Education of head and If7 years or older: What is hidher If6 years or older: What If6 years or older: What max in hh highest level of education is hidher highest level of is hisher highest level of completed? educationcompleted? educationcompleted? 8 categories 8 categories 8 categories Age (of householdhead) How old is heher?Age at last How old is heher? Age at How old is heher?Age at birthday. Note "0" for children last birthday. Note "0" last birthday. Note ``0" under 1year for children under 1year for children under 1year Sex of householdhead What is hisher sex? What is hisher sex? What is hisher sex? StudentAJnemployment What is hisher main activity What is hisher main What is hisher main status duringthe past 12 months? activity during the past 12 activity during the past 12 11categories months? months? 11 categories 11 categories Income source What is the main source of income What is the main source What is the main source for this household? of income for this of income for this 8 categories household? household? 8 categories 10 categories Housing: brick walls What is the main material of the Does your dwelling have Does your dwelling have walls? 4 categories; walls of block, bricks or walls of block, bricks or 2=bricks/cement blocks. cement? cement? Housing: electricity I s this dwelling electrified? Does your dwelling have Does your dwelling have electricity? electricity? Housing:tap water What is the main source of water Does your dwelling have Does your dwelling have for this household? Central indoor water taps? indoor water taps? pipeline, own pipeline, standing water pipe. Purchaseof durables Has anyone in the household Has anyone inthe Has anyone in the duringthe last 12 months householdduringthe last householdduring the last purchased any ...? any...? 12 months purchased 12 monthspurchased 57 categories any...? 57 categories 57 categories Source: World Bank staff estimates from HBS data. 31 Table A.6: Definition of ConsumptionAggregates for the Different Methodologies PA 05 CA Revised* ComparableCA Foodexcluding own production 4 4 4 Alcohol and tobacco 4 4 In-kind(received) 4 4 Own production 4 4 4 Non-foodexcl health and education 4 4 Education 4 4 Health 4 Utilities excl value of housing 4 4 Value of housing Source: World Bank staffestimatesfrom HBS data. Notes: Certainhighvolatility items are excluded (air and sea * transportationexpenses; gamblingand holiday packages;financial andjudicial services). Utilities includedomestic services. Table A.7: PovertyLinesin Different Methodologies(in Euros, per adult equivalent per month) Methodology 2002 PL Robust Endogenousin Endogenous Endogenousin adjusted in Comparable in Econometric with CPI Comparable Surveys Econometric Projection CA Projection (unweighted) (weighted) Wave Povertvline 2002103 43.12 45.6 44.1 2003104 43.10 41.8 37.2 45.6 44.0 2004105 43.34 44.2 38.1 46.8 45.2 2005106 43.10 40.39 38.1 46.8 45.2 Source: World Bank staff estimatesfrom HBSdata. 32 ANNEX B: RESULTSUSINGDIFFERENTSURVEY YEAR DEFINITION 3.8 In this Annex, we present our results using survey waves defined by the introductionof changes inthe questionnaire. 3.9 There is a differencebetween the sampling timing and the timing of the introductionof changes in the questionnaire. The sampling is done for the household being survey in June through May each year. Changes and additions to the survey questionnaire are introduced in January, starting 2005. The selection of which households and EAs are sampled each month is not clear, although 200 households from 25 EAs are consistently surveyed. There is evidence, however, that the surveying consequence is not representative by month or half a year. For instance, much largershare ofthe populationis surveyedeach second half of the year than during the rest of the survey. Partitioningthe sample by calendar year, thus, introduces a bias. Indeed, the resultsusingwaves defined as inthe table belowshow a different trend inpoverty. Table B.l: Introduction of New Questionnaires KOSOVOHBS Wave I Wave I1 Wave111 Wave IV HBS-2002-2003 HBS-2003-2004 HBS-2004 HBS-2005 Period 612002-512003 612003-612004 612004-1212004 112005-1212005 Numberof 2400 households 2400 households 1400 households 2400 households observations (960 rural, 1440 urban) Table B.2: Poverty Statistics using PA05 Methodolopy Absolute Poverty Headcount Extreme Poverty Headcount Weighted Unweighted Weighted Unweighted 612002-512003 37.93 43.56 15.43 18.30 612003-612004 45.14 41.83 13.85 13.64 612004-1212004 35.79 31.61 12.43 10.39 112005-1212005 39.72 39.13 12.68 13.24 Source: World Bank staffestimatesfrom HBS data. Note: Unweightedhere refers to no weights being used andthus these estimatesare at household-levelversus the population-levelestimatesinthe "weighted" column. 33 Table B.3: Poverty Rates Using PA05 Methodology Absolute and extreme poverty rates Urban and Rural Poverty Rates Urban Rural 2002103 46.99 34.49 2003104 42.73 46.01 06-1212004 30.13 38.73 2005 34.95 42.39 ...-. ... _.........................-* _ _ _ _ .. I.. %wo. HBS I, II.111and IV PovmyA-~nant(2005) rnebdolowyuwd Source: World Bank staff calculations from HBS data Table B.4: Detailed Poverty Diagnosticswith Revised Consumption Aggregate Absolute Poverty Headcount Extreme Poverty Headcount By wave weighted unweighted weighted unweighted 612002-512003 40.6 46.0 '17.9 20.1 612003-612004 46.9 42.8 14.5 14.4 612004-1212004 37.3 33.0 12.8 10.7 112005-1212005 42.1 40.7 13.1 13.7 Source: World Bank staffcalculationsfrom HBS data. Unweightedhererefers to no weights beingused and thus these estimatesare at household-levelversus the population-levelestimatesin the "weighted" column. 34 Table B.5: PovertyRatesUsingAlternative Consumptionand Poverty LineMethodologies ConsumptionAggregate (CA) Povertv Rates definition CAL CA withPA05 PL 2002 adjustedwith CPI methodology weighted unweighted By wave 2002-03 37.9 43.6 2003-04 45.1 41.8 2004-05 35.8 31.6 2005-06 39.7 39.1 CA IL ComparableCA (Lanjouw and Lanjouw, 2001) PL 2002 adjustedwith Nonparametric Poverty CPI Line weighted unweighted weighted unweighted By wave 2002-03 40.6 46.0 2003-04 46.9 42.8 39.6 35.9 2004-05 37.3 33.0 39.2 34.8 2005-06 42.1 40.7 38.6 37.5 CA I1in ComparableSurveys PL 2002 adjustedwith EndogenousPoverty Line CPI weighted unweighted weighted unweighted By wave 2002-03 40.6 46.0 2003-04 46.9 42.8 37.8 34.9 2004-05 37.3 33.0 29.9 25.7 2005-06 42.1 40.7 33.2 32.5 CA IUsingEconometricPoverty Projection (Stifel and Christiansen, 2006; Poverty Line Elbers, Lanjouw, and Lanjouw, PL 2002 adjustedwith EndogenouslyDetermined 2003) CPI By wave weighted unweighted weighted unweighted 2002-03 29.7 34.7 38.2 36.2 2003-04 24.7 31.1 30.2 31.1 2004-05 28.8 30.1 38.8 32.3 2005-06 26.8 28.7 35.5 30.7 Source: World Bank staff calculations from HBS data. 35 REFERENCES AMP Kosovo. 2006. The Rural Development Context of Kosovo. Agricultural and Rural Development Planfor Kosovo. (downloaded from http://www.ampkosovo.corn/ardpl .htmlon 9/12/2006). Carletto, C. ,Davis, B. and Stampini, M.,Zezza, A. (2004). A Country on the Move: Internal Mobility and InternationalMigration in Post-Communist Albania. FA0 (mimeo) Deaton, A. and Zaidi, S. 2002. Guidelines for Constructing Consumption Aggregates for Welfare Analysis. LSMS WorkingPaper 135. Deaton, A. and Dreze, J. 2002. Poverty and Inequality inIndia: A Re-Examination. Economic and Political Weekly September 7,2002 Deaton, A. 2003. Adjusted IndianPoverty Estimates for 1999-2000. Economic and Political Weekly January 25,2003 Demery, L. and Groootaert, C. 1993. Correcting for Sampling Bias inthe Measurement of Welfare and Poverty inthe Cote D'Ivoire Living StandardsSurvey. The WorldBank Economic Review. Vol 77No 2: 263-292. Elbers, C., Lanjouw, J., Lanjouw, P.,. 2003. Micro-levelEstimation of Poverty and Inequality. Econometrica, Vol. 71, No 1(January, 2003), 255-364. Gibson, John. 2005. Statistical Tools and Estimation Methods for Poverty Measures Basedon Cross-SectionalHousehold Surveys. Handbook on Poverty Statistics: Concepts,Methods and Policy Use, 128-205 (downloaded from http://unstats.un.or~/unsd/methods/povert~/~df/Chapter-5on 9/12/2006) .pdf Grimm, M. and Gunther, I.2004. How to achieve pro-poor growth ina poor economy: The case of BurkinaFaso. Operationalizing Pro-Poor Growth" (downloaded from " www.wor1dbank.org on 9/12/2006), Howes, S. and Lanjouw, J. 1997. Poverty Comparisons and Household Survey Design. LSMSWorking Paper Number 129. Lanjouw, P., and Kijima, Y. 2003. Poverty inIndiaduring the 1990s:A Regional Perspective. World Bank Policy Research Working Paper 3141. Lanjouw, J, and Lanjouw, P. 2001. "How to CompareApples and Oranges: Poverty MeasurementBasedon Different Definitions of Consumption" Review of Income and Wealth Series 47, Number 1, March2001. 37 Lohr, S. 1999. Sampling - Design and'tlnalysis. 1st Edition, Arizona State University, 0534353614. Lynn, P. 2004. "Development ofa Sampling Methodfor Household SurveysinPost- War Bosnia and Herzegovina" Statistics in Transition June 2004 Vol. 6, No. 6, pp. 953-977 Ravallion, M. 1994. "A Better way to set poverty lines, "Outreach #15, March 1994. Statistical Office of Kosovo (SOK). 2005. Agricultural and Household Survey 2005. Series-Agricultural and Environment Statistics, Pristina, Kosovo. Stifell, D. and Christiaensen, L. 2006. Tracking Poverty Over Time inthe Absence of Comparable Consumption Data WPS3810. Tarozzi, A. 2005. Calculating Comparable Statistics from Incomparable Surveys, with an Application to Poverty inIndia (downloaded from http://www.econ.duke.edu/-taroz/TarozziComparabilitv.pdf on 9/12/2006) 38