Using Survey Data to Improve Registry Data The Case of Syrian Refugees Paolo Verme World Bank
Motivation UNHCR keeps records of all refugees registered with the proGres database UNHCR collects a rich set of additional information via home visits’ surveys This information is little exploited for analysis UNHCR needs to target refugees with its cash assistance program due to funding shortages How good is the UNHCR targeting capacity? Can UNHCR improve its targeting capacity by making better use of available information and reduce the cost of surveys?
Background A WB-UNHCR partnership: WB analytical expertise on poverty and welfare and UNHCR expertise on refugees Pilot study in March 2014 Two countries (Jordan and Lebanon) 6 Data sets (3 Jordan and 3 Lebanon): UNHCR registry for Jordan and Lebanon (PG: 650,000 records Jordan; 1.2 m records Lebanon); UNHCR and WFP home visits (45,000 cases in Jordan, 2 rounds; 30,000 cases Lebanon, 1 round) UNHCR and WFP survey (1,700 cases Lebanon, 1 round) Focus on refugees living outside camps How Poor Are Refugees? A Welfare Assessment of Syrian Refugees Living in Jordan and Lebanon (16 December, 2015)
Basic Idea 1.Construct welfare aggregates 2.Construct a welfare and poverty model with PG+HV data 3.Tests best proxies for welfare and poverty 4.Tests composite indexes for welfare 5.Derive the best welfare and poverty models (PG+HV data) 6.Find the best variables to add to the proGres data 7.Predict welfare and poverty using PG data 8.Use the new PG model for targeting assistance programs
1) Construct Welfare Aggregate
2) Construct Welfare Models
3) Test best proxies of Welfare Basic tabulations Univariate regressions on 185 variables Ranking by R2 Multivariate tests: Backward and forward regressions Manual tests of key explanatory variables Tests repeated for composite welfare indicators Maximizing R squared
3) Test best proxies of Welfare
VariableDescriptionObsW R2 lin csize_hvIndividuals in case (HV) dem_p_childProportion of children edu_p_attendProportion of children in school pov_inc_unhcrUNHCR Monthly Financial Assistance cash_large Large Family &/or Family with Babies, Toddlers or Children Attending School prot_bail_date pov_cop_aidHumanitarian assistance cash_smotherSingle Women prot_aggrScore for work and residence permit, MOI and bail out doc cash_elderlyElderly 60 and Above prot_moi_diff_apoliceNot comfortable approaching a police station pov_inc_aggrTotal number of kinds of income prot_bailBailedOutFromCamp cash_decisionDecision_for_cash_Assistance pov_cop_hostLiving together with host family ref_elderlyElderly alone prot_know_schoolSchool edu_p_notattendProportion of children not in school house_sanitarySanitaryFacilitiesStatus dem_pafemaleFemale Principal Applicant
4) Test Composite Indexes of Welfare VariableObsMeanStd. Dev.MinMaxR2% i_rent i_latrine i_good_liv~d i_housecon~n i_pipewater i_good_san~y i_good_ven~n i_waste i_water i_good_ele~y
4) Test Composite Indexes of Welfare VariableObsMeanStd. Dev.MinMaxR2 ind_house_crowd ind_house_crowd ind_wash_water ind_nfi ind_house_subjective ind_house_assets ind_cope_index ind_wash_hygiene ind_cope_wfp ind_food_wfp ind_house_quality ind_food_score ind_food_variety
5) Derive the optimal Welfare Model (PG+HV )
Welfare Model (Cont.)
6) Select the Best Variables to Add to PG Data The poverty model predicts poverty correctly 90.1% of the times PG-HV predictors: Case size Rent Place of destination (country and region) Official entry and point of entry Place of origin (Damascus vs other regions) HV predictors (h) Principal applicant characteristics (age and marital status) Selected assets (latrine, piped water, kitchen) Education and former occupation are less important than expected
7) Predict Welfare using new PG model Wexp_unhcr_lncap Coeft Individuals in case (HV) *** Proportion of children *** Concrete House 0.195***8.017 Santitation average or above 0.109***7.244 Ventilation average or above 0.100***6.194 Free Housing *** Proportion school-aged children 0.113***8.919 Proportion of children in school *** Sharing costs with host family *** Living together with host family 0.114***8.987 IsCertificateValid 0.124***9.190 _cons 4.715*** Number of observations 14,150 R Adjusted/Pseudo R
8) Use New PG Model to Target Assistance Income ExpenditureNon poorPoorTotal Non poor Poor Total
Conclusion World Bank and UNHCR have complementary skills and resources Existing UNHCR data can be used to improve targeting: Shift from income to expenditure Reduce leaking and costs Improve coverage and targeting The analysis highlighted areas to improve: Data collection (survey design, sampling, questionnaires) Data management (PG management system, survey administration) Data analysis (economists and social protection officers)