Policy Research Working Paper 10093 Bridging the Targeting Gap Assessing Humanitarian Beneficiaries’ Likely Eligibility for Social Protection in Iraq Lokendra Phadera Dhiraj Sharma Matthew Wai-Poi Lotti Douglas Vladimir Jovanovic Oliver Westerman Safwan Aziz Khan Poverty and Equity Global Practice June 2022 Policy Research Working Paper 10093 Abstract In Iraq, the Multi-Purpose Cash Assistance programs have and the humanitarian Multi-Purpose Cash Assistance pro- been instrumental in reaching the households most affected grams. Using the common proxies between the cash transfer by the conflict with the Islamic State of Iraq and the Levant targeting formulas of the humanitarian agencies and the in areas where the coverage of the government’s social safety government, the pseudo-proxy-means test provides each net (SSN) programs remain limited. In the evolving context, Multi-Purpose Cash Assistance beneficiary’s probability of however, short-term Multi-Purpose Cash Assistance pro- being eligible for the government’s cash transfer program grams will require eventual integration in some form with under different expansion scenarios. When applied to the the government’s social safety net programs to continue existing humanitarian beneficiary database, the results of reaching vulnerable households affected by the conflict. the pseudo-proxy-means test tool suggest the potential for As an initial step, this paper proposes an analytical pseu- both significant referral numbers and a sequenced referral do-proxy-means test (PPMT) tool to bridge the targeting strategy. differences between the government’s cash transfer program This paper is a product of the Poverty and Equity Global Practice. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The authors may be contacted at lphadera@worldbank.org; dsharma5@worldbank.org; mwaipoi@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Bridging the Targeting Gap: Assessing Humanitarian Beneficiaries’ Likely Eligibility for Social Protection in Iraq 1 Lokendra Phadera, Dhiraj Sharma, & Matthew Wai-Poi World Bank Lotti Douglas & Vladimir Jovanovic Mercy Corps Oliver Westerman CLCI / Oxfam Safwan Aziz Khan CWG / UNHCR Keywords: Social Protection and Growth; Proxy Means Tests; Targeting in Social Protection; Humanitarian Assistance; Cash Transfer; Internally Displaced Persons; Iraq JEL Classification: D04, H53, I32, I38 1 Acknowledgment: We acknowledge the contributions of the UK government’s Foreign, Commonwealth and Development Office (FCDO), which convened the initial workshop and launched this collaborative effort. This work was supported by the Iraq Reform, Recovery, and Reconstruction Fund (I3RF) trust fund. We are grateful for the guidance of the two peer reviewers Sharad Alan Tandon and Phillippe George Leite, and for the continuous support and feedbacks from Rene Antonio Leon Solano, Gabrielle Fox, Virginia Leape, Maria Ana Lugo, and the World Bank’s Iraq Social Protection team. Our thanks are also due to the Central Statistics Office (CSO) and the Kurdistan Regional Statistics Office (KRSO) for data sharing and collaboration. This work was undertaken under the overall guidance of Ramzi Afifi Neman, Johannes G. Hoogeveen, and Benu Bidani. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the government they represent. 1 1. Introduction The humanitarian situation in Iraq is in a transition phase, evolving from the response to the conflict with the Islamic State of Iraq and the Levant (ISIL) to a humanitarian context with unforeseen protracted displacement and delayed implementation of recovery and resilience activities. Since January 2014, over 6 million (about 15 percent of the population) got displaced due to the conflict (IOM Iraq, 2020a). While 78 percent of those who fled have returned to their areas of origin (IOM Iraq, 2020b), those who were among the initial wave of returnees have in some cases found destroyed or damaged houses, unexploded ordinances, absence of livelihoods and services, lack of social cohesion (community tensions and fears of population change) and security concerns (blocked return or perceived insecurity). Cash Transfer Programming (CTP) has been used in Iraq since 2014 to support conflict-affected and displaced populations, initially for Syrian refugees living in camp and out-of-camp settings across northern Iraq and the Kurdistan region of Iraq (KRI). In late-2014 and 2015, the internal displacement following the conflict with ISIS led to cash and voucher programs being implemented at a larger scale by NGOs, UNHCR, and WFP. The Cash and Livelihoods Consortium for Iraq (CLCI, formerly CCI) was formed in early 2015 as a partnership between the largest international NGO CTP actors to develop harmonized tools and approaches, and a cohesive MPCA strategy for Iraq. 2 The Iraq Cash Working Group (CWG) was initially formed in 2014 as a ‘semi-cluster’ to coordinate the growing Multi-Purpose Cash Assistance (MPCA) response and was subsequently given a seat at the Inter-Cluster Coordination Group (ICCG) and, following advocacy by MPCA actors and the CLCI Steering Committee, developing a standalone MPCA chapter in the 2015 Humanitarian Response Plan (HRP). MPCA has had a chapter in the Iraq HRP every year since. Over a million conflict-affected households have been reached with MPCA since the start of the conflict with ISIS. While the Ministry of Migration and Displaced (MOMD) and humanitarian agencies are providing temporary support to affected populations, access to relevant social protection programs remains challenging for both internally displaced persons (IDPs) and returnees. Obstacles include a lack of necessary civil documentation for enrollment in social programs, inadequate assessment capacity of the authorities and complicated registration procedures. A responsible transition of the humanitarian cases towards programs financed by the government and supported by international development planning frameworks, e.g., the Rapid Recovery Program (RRP), Funding Facility Stabilization (FFS), United Nations Development Assistance Framework (UNDAF) and the United Nations Sustainable Development Cooperation Framework (UNSDCF)), is urgently needed. 3 Recognizing the need for a responsible transition of the humanitarian cases to the national social protection system, a conversation was initiated in April 2018 among the Government of Iraq, humanitarian agencies, bilateral partners, and multilateral institutions. The early discussions led to a two- day Social Protection workshop convened by the UK government’s Foreign, Commonwealth and Development Office (FCDO), formally DFID, with the Ministry of Labor and Social Affairs (MoLSA), the World Bank (WB), the Cash and Livelihoods Consortium for Iraq (CLCI), the CWG and key UN agencies in attendance. It was recognized that MoLSA’s Social Protection Network (SPN), which reaches 1.36 million 2 https://www.calpnetwork.org/wp-content/uploads/2020/03/calp-inter-agency-collaboration-web-1.pdf 3 See Annex C for a more complete description of both formal government social protection programs and humanitarian assistance currently in Iraq. 2 Iraqi households through its targeted cash transfers, was the best placed institution to meet the needs of conflict-affected households vulnerable to future shocks or poverty. 4 One of the main outcomes of the workshop was the agreement to undertake a joint desk review by members of the CLCI, CWG, and the WB to identify the potential for closer alignment between humanitarian cash assistance and MoLSA cash transfer program, as well as possible transitional pathways from the former to the latter. To do so, one of the first tasks was to gauge what share of humanitarian caseloads would be eligible for MoLSA’s cash transfer program were they to be referred to the MoLSA program. Although the proxies used in the models were different, both the national and humanitarian CT programs used proxy-means-test (PMT) method to assess eligibility. Had all the information been available to apply the national PMT on the CLCI’s database, this step would have been straightforward. But it was not possible to evaluate MoLSA’s PMT on the CLCI’s current caseload because not all the indicators necessary for the national PMT were available in the CLCI database. To overcome the impediment, we develop an analytical tool called pseudo-PMT (PPMT) that provides a bridge between the targeting models of the two CT programs for potential beneficiaries. Like a regular PMT, the PPMT relies on proxies and their weights to calculate potential beneficiaries’ well-being score (PPMT-score) and uses it to identify program eligibility. Unlike in a regular PMT where an optimum set of covariates that correlates best with the household welfare variable (consumption, income, etc.) are selected from an extensive list, the PPMT uses only proxies that are common to both the humanitarian’s and the government’s CTP targeting formulas. By design the PPMT cannot reproduce the score that would be produced by the full PMT model. Therefore, we can only assess the probability that a case with a certain PPMT score will be eligible for the MoLSA’s program determined using the full PMT model. Since both the PPMT and the full PMT for MoLSA’s program are devised using the same national household survey, the 2017/18 Survey of Well- Being via Instant and Frequent Tracking (SWIFT), we use the data to calculate the probability of qualification under the full PMT given a household’s PPMT score. Households with lower PPMT score will have higher probability of being eligible under the full PMT and vice-versa. One can be conservative in making the referral recommendation, i.e., referring a caseload to MoLSA only if its PPMT score is very low so that its probability of selection for the MoLSA program is very high. The benefit of doing so is a high “hit” rate, but it may be costly in that many cases that would be eligible are not referred. On the other hand, one could be liberal in making the referral recommendation by increasing the PPMT score threshold below which cases will be recommended to MoLSA’s review. This reduces the chances that potentially eligible cases do not get an opportunity for MoLSA’s screening because they are not referred. However, it will add significant administrative burden on MoLSA as many cases will be referred for verification. Results from applying the pseudo-PMT models to the CLCI’s legacy database of current and past beneficiaries (37,457 households corresponding to 224,736 individual) suggest potential for both significant referral numbers and a sequenced referral strategy. 5 The analysis suggests that referring only those in the bottom 18 percent nationally (the national poverty rate according to SWIFT 2017/18 survey 4 Throughout the paper, Social safety net (SSN), Social Protection Network (SPN), national program, MoLSA’s cash transfer program etc. are used loosely and interchangeably to refer to the government’s cash transfer program delivered by the Ministry of Social and Labour Affairs (MoLSA). 5 Due to the CLCI’s CT program coverage, the database is limited to the Northern or ‘conflict-affected’ governorates: Anbar, Diyala, Kirkuk, Ninewa, Salah al-Din. 3 without adjusting for missing districts), 24 percent of the existing humanitarian database would be eligible for referral to the MoLSA’s CT program with a high confidence level (90 percent or more). The referral rate increase with the lower referral confidence, about 54 percent of the humanitarian caseload would be eligible for the referral under at a low confidence level (50 percent or higher). To allow flexibility depending on the availability of budget, the analysis is also carried out by expanding the coverage of the national program and changing the referral confidence rate. In general, the number of referral cases increase with higher coverage and lower referral confidence. Given a large proportion of the humanitarian database that could potentially be referred for MoLSA CT program, the results suggest a sequencing of smaller numbers with a high degree of likelihood to qualify could be referred first, with larger numbers being referred later when budgets expand. The remainder of the paper is structured as follows. Objectives of the joint activities are outlined in section 2. Section 3 presents the analytical methodology, including the proposed pseudo-PMT (PPMT), its performance criteria, and theoretical results. Cross-eligibility results from applying the PPMT on the legacy database and potential eligible households’ profile are presented in section 4. Concluding remarks and discussions of next steps are outlined in section 5. 2. Objectives and Activities The main objective of this initial desk analysis is to understand the extent to which beneficiaries of humanitarian cash assistance would be eligible for MoLSA social protection programs under the MoLSA targeting approach. This means both understanding: 1. The percentage of the existing humanitarian database of current and past beneficiaries identified under the current vulnerability assessment approach who are likely to be eligible; 2. The percentage of new beneficiaries who would be identified under the new vulnerability assessment approach who would be likely to be eligible. In addition, it is desirable to understand who would be eligible and who would not. This requires understanding where eligible (and ineligible) households live, as well as their demographic characteristics. In order to meet these objectives, three activities have been undertaken: 1. An analysis has been conducted of the targeting overlaps between the MoLSA approach and the historical humanitarian approach. 2. A similar analysis has been conducted of the MoLSA approach and the new humanitarian approach which had recently been developed. 3. Geographic and demographic profiles of the likely eligible (and likely ineligible) are analyzed. 3. Methodology The challenge in assessing the likely percentage of beneficiaries of humanitarian cash assistance who could be referred to the SSN program is that the humanitarian database lacks all the variables necessary 4 to implement the national PMT. Initial discussions between the government, development partners and humanitarian actors revealed a division in terms of concepts, language and philosophy. While the humanitarian programs use vulnerability, the SSN uses monetary poverty for eligibility assessment and the targeting models between the two differ. However, the technical work summarized in this note allowed initial work to begin and momentum and discussion to ensue. Genuine concerns over data sharing issues were able to be overcome through the sharing not of underlying data but of pseudo-PMT models which captured the commonalities of targeting approaches without running afoul of data privacy. This section outlines the MoLSA eligibility criteria and compares it to the current humanitarian approach. It then proposes a PPMT tool to compare these two methods given the lack of common data set. 3.1 MoLSA Eligibility According to the new 2014 Social Protection Law, only households below the poverty line are eligible for social assistance support. The poverty line in 2017 was IQD 110,881 per person per month (Sharma & Wai- Poi, 2019). By this line, national poverty was 20 percent in 2017-18 (Sharma & Wai-Poi, 2019), 6 slightly above the 18.9 percent measured directly in 2012 (World Bank, 2014) but a decline from 22.5 percent in 2014, and 41.2 percent in affected governorates, which was estimated indirectly by simulation methods (Krishnan & Olivieri, 2016). Thus, the current 1.36 million SPN beneficiaries (IMF, 2021) are less than the estimated number of poor in Iraq, highlighting the legal (but not fiscal) space for expanding the program. A full revision of the poverty line and poverty rate will be done after the implementation of the third round of the Iraq Household Socioeconomic Survey (IHSES). In practice, MoLSA does not observe potential household beneficiaries’ poverty status, as this would require administering an intensive, time-consuming and expensive household expenditure survey which is not practical to be conducted for all potentially eligible households. Instead, the MoLSA determination of eligibility is based on the PMT scoring system and the application process has the following steps: 1. Application: a. Apply on the MOLSA website, screening questions determine if a home visit is warranted; b. Applications process closed May 2016, with 340,000 waiting to go for MOP scoring. 2. Home visits: a. For those households passing the screening question, social workers visit the home and collect data on a range of indicators as advised by the PMT formula. 3. Household scoring: a. Ministry of Planning (MOP) estimates household per capita consumption using the PMT indicators; b. Models updated when new national household survey data available. 4. Eligibility determined: a. Households with estimated per capita consumption under the poverty line. 6 The 2017/18 Survey of Well-Being via Instant and Frequent Tracking (SWIFT) could not be implemented in 14 of the 120 districts due to continuous insecurity. The direct poverty estimate from the data without simulating for the missing districts is 18 percent. 5 5. Validation: a. Households cross-checked with pension and land cadaster data to confirm they are not wealthy. 6. Entry or waitlisting: a. Households enter program if sufficient budget, otherwise go on the waitlist. 7. Reverification: a. By law, households file an annual report on the MOLSA website; b. Social workers also revisit the households each year to maintain rapport and support case management. The key to MoLSA eligibility is the PMT model. This model uses welfare proxies to predict consumption. Such proxies need to be generally easy to verify and correlated with consumption. Common categories include: (i) demographic; (ii) education and employment; (iii) housing; (iv) assets; and (v) displacement. Statistical models are developed giving scoring weights to each proxy using national household surveys. The initial models were based on the 2012 IHSES survey. New models have been developed using the more recent 2017/18 Rapid Welfare Monitoring Survey, also known as the Survey of Well-being via Instant and Frequent Tracking (SWIFT) and are currently being implemented. Each household has an estimated consumption based on these weights. 3.2 MPCA Eligibility The CLCI used a similar but distinct approach to develop a tool for MPCA vulnerability assessment and scoring, endorsed by the CWG and widely used by other MPCA actors including I/NGOs as well as UN agencies. However, the 2016 Vulnerability Model (VM) differs in a number of ways: 1. Both MoLSA PMT and VM use statistical regressions to derive weights for each indicator, but vary in what they are trying to estimate: a. PMT estimates household per capita consumption; b. VM estimates income to earnings ratio (initially estimated $2 a day poverty status). 2. They vary in how they do this: a. PMT uses a continuous regression with variable selection (stepwise in 2012, Lasso in 2017); b. VM uses a binary regression with variable selection (stepwise). 3. They vary in the number of geographical models: a. PMT produces regional models; b. VM produces a single model for all regions. The CLCI revised the models using the 2018 Multi Cluster Needs Assessment (MCNA) survey, which resulted in the 2019 Socio-Economic Vulnerability Assessment Tool (SEVAT).7 The new models are closer 7 https://www.humanitarianresponse.info/en/operations/iraq/document/iraq-mpca-vulnerability-model-review 6 to the MoLSA PMT models, in that they both use per capita consumption as a dependent variable and they both have regional models. The current work compares the potential overlap between MoLSA PMT and both old (VM) and new (SEVAT) MCNA models. The next section describes how this is done. 3.3 Pseudo-Proxy Means Test (PPMT) Methodology The 2016 Vulnerability Assessment data 8 and the 2018 Multi Cluster Needs Assessment (MCNA) survey that formed the basis for CLCI’s VM and the SEVAT models do not have the full set of SSN PMT variables. To fill this assessment gap, we propose a probabilistic pseudo-PMT (PPMT) formula that includes only indicators that are common to the national PMT and the vulnerability models (VM and SEVAT) and provides its performance criteria. Using the data from the nationally representative SWIFT survey, we run an ordinary least squares (OLS) regression of household consumption on the common set of proxy indicators, Xs. For each common indicator, we assign a pseudo-weight based on its estimated association to household consumption as below: = + ′ + (1) where, is a per capita monthly expenditure of household in natural log, which is a function of a vector of covariates/proxies, , and a random idiosyncratic error term, . ̂ s are the estimated weights � and for a constant term and pseudo-proxy variables respectively. As discussed in Box 1, we acknowledge the possibility of alternative approaches than the one presented here. There are, however, genuine concerns over data privacy and sharing, divisions in terms of concepts, and philosophy. Given the circumstances and scope of preliminary focus to be purely technical, we believe the tool could prove useful and is technically robust to allow the initial work to begin and momentum and discussion to ensue. Box 1: Possible alternative approaches? – A discussion The method employed in the paper may not be the only approach possible in this context. Creating a PPMT that is limited to the overlapping proxies between the two PMTs may be restrictive in terms of number of available variables for model prediction. An alternative could be, rather than focusing the comparison between the two targeting models, comparing available information in the underlaying data sets of the two models may be more flexible. Potentially more variables will overlap between the data sets than between models. Then following the procedure described in the paper, one would estimate weights using the household survey that was used to devise the national PMT and create the probabilistic referral thresholds. One will even have an option to select an optimal set of variables and strike the balance between parsimony and goodness of fit when there are too may overlapping variables. With potentially more and/or better correlated covariates, the new pseudo-PMT may have more accurate referrals than the proposed PPMT when applied to the humanitarian database. 8 Data collected across several CCI and other CWG partners were harmonized for the 2016 “Beneficiary Vulnerability Analysis” report. The harmonized data formed a basis for the 2016 VM model. See Cash Consortium of Iraq (2016). 7 In practice, this would require knowing the type of information available in both the data sets. The questionnaire of the household survey, which forms the basis for the national PMT, is readily available, but what indicators can be constructed using the humanitarian database in not public information. Although, there is no need for sharing the micro-data, concerns related to data sharing and running into data privacy infringement are a possibility. How many MoLSA beneficiaries could come from the humanitarian database will depend on circumstances. The displaced households, by all descriptions, are likely to be severely vulnerable. Thus, depending on the availability of budget, political will, and the technical and administrative capacity, referral may, even, rely on non-PMT approaches such as group/geographical targeting or combination of the two. Moreover, perceived ISIS affiliation, fear of persecution and other grievances may mean not all eligible households will be willing to be referred. The possible transitional pathways, thus, will differ depending on these circumstances. Following the completion of the MCNA survey in 2018, humanitarian agencies updated the vulnerability assessment criteria for CLCI program with separate models for each region, namely: Kurdistan (comprising three governorates in Kurdistan), North (comprising five Northern governorates), and rest of the country (Center and South) corresponding to the World Bank’s regional definition for PMT formula. To be consistent with the new targeting and to assess eligibility into the government’s CT program (SSN) of both the previous and the potential future CLCI beneficiaries we formulate two sets of regional pseudo-PMTs: 1. Three regional models based on common set of indicators between the SWIFT and the 2016 VM model that was devised on the Vulnerability Assessment data (hereafter VM); 2. Three regional models based on common set of indicators from the SWIFT and the updated VM model (SEVAT) that was devised on the 2018 MCNA survey (hereafter SEVAT). Means and standard deviations of the variables common to the national model and the old VM model are reported in Table 1. A single national VM model was devised in 2016, hence, the variables for all three regional models are the same. In total, 5 variables are in common between the two sources: household employment rate, dependency ratio, indicator for female headed household, household size, and a dummy to indicate households with IDP member/s. Table 2 reports the indicators selected for the new CLCI eligibility criteria, SEVAT, that were also considered for the SSN formula. Only 3 variables, home ownership, household size and employment rate, from the new Kurdistan and Center-South MCNA models can be duplicated in the national model. However, an additional variable, access to public network water tap, can be reproduced from the northern SEVAT model. Table 1. Common variables between the SWIFT and the 2016 VM model (1) (2) (3) (4) Iraq Kurdistan North Center- South Mean [SD] Mean [SD] Mean [SD] Mean [SD] Employment rate (emp/hhsize) 0.22 0.21 0.21 0.23 [0.15] [0.15] [0.14] [0.15] Dependency ratio 0.42 0.40 0.43 0.43 8 [0.22] [0.23] [0.22] [0.21] Female headed household 0.09 0.05 0.10 0.10 [0.29] [0.22] [0.30] [0.29] Household size 7.40 6.48 6.99 7.83 [3.60] [2.53] [3.27] [3.91] IDP 0.07 0.12 0.14 0.02 [0.25] [0.32] [0.35] [0.12] Observations 1500 540 300 660 Note: Standard deviations are in brackets. Data source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. Table 2. Common variables between the SWIFT and the SEVAT model (1) (2) (3) (4) Iraq Kurdistan North Center-South Mean [SD] Mean [SD] Mean [SD] Mean [SD] Home: not owned by HH 0.27 0.30 0.29 0.26 [0.45] [0.46] [0.45] [0.44] Household size 7.40 6.48 6.99 7.83 [3.60] [2.53] [3.27] [3.91] Employment rate (emp/hhsize) 0.22 0.21 0.21 0.23 [0.15] [0.15] [0.14] [0.15] Water: no public network tap 0.08 0.05 [0.27] [0.22] Observations 1500 540 300 660 Note: Standard deviations are in brackets. Data source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. The PPMT formulae for each region corresponding to the 2016 VM model and the new SEVAT models are reported in Table 3 and Table 4 respectively. The coefficients, weights, are estimated using equation 1. Given the small set of covariates, it is not surprising that the R-squared for both sets of models are relatively low - below 30% for both the Kurdistan and Center-South models and about 50% for the North. While employment rate is positively correlated, all the other variables in the old models are negatively correlated with log per capita consumption apart from the female headed household and IDP variables in the Center-South model, which are positively correlated (Table 3). In the new models, household size and not owning a house have a negative association with consumption; employment variable, in contrast, has a positive association. Surprisingly, not having access to public network appear to be a positive predictor of a household’s measure of welfare (Table 4). These pseudo coefficients are applied on the appropriate beneficiary registration database to understand what proportion of the existing humanitarian caseload would be eligible for the government’s cash transfer program. Table 3. Pseudo-PMT for the 2016 VM model (1) (2) (3) Kurdistan North Center- South 9 Employment rate (emp/hhsize) 0.20 0.55 0.72 Dependency ratio -0.49 -0.44 -0.31 Female headed household -0.09 -0.03 0.10 Household size -0.07 -0.09 -0.05 IDP -0.01 -0.28 0.08 Constant 5.98 5.84 5.56 Observations 540 300 660 R-squared 0.24 0.50 0.28 Data source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. Table 4. Pseudo-PMT for the SEVAT model (1) (2) (3) Kurdistan North Center-South Home: not owned by HH -0.03 -0.14 -0.04 Water: no public network tap -- 0.12 -- Household size -0.07 -0.10 -0.05 Employment rate (emp/hhsize) 0.48 0.83 0.93 Constant 5.70 5.65 5.40 Observations 540 300 660 R-squared 0.19 0.47 0.27 Data source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. 3.4 Cross-Eligibility Calculation and Referral Accuracy This exercise attempts to create a tool to assess the eligibility of the CLCI’s beneficiaries for the MoLSA’s CTP. Given the small set of covariates that overlap between the two eligibility criteria, the pseudo-PMT cannot solely determine qualification into the MoLSA’s CTP.9 However, it allows to evaluate the eligibility 9 Although all pseudo-PMT indicators were included in the set of candidate proxies, the final MoLSA PMT may not reflect some. A Lasso statistical technique was used as a model selection device for the development of the full PMT. Out of numerous candidate proxies, Lasso selects an optimal set of covariates that jointly have the best predictive power. Coefficients of the proxies with lesser predictive power are zeroed-out and not selected. See 10 probabilistically. That is, given a pseudo-PMT score, , of a household we can calculate its probability of qualifying for SSN under the full PMT. 10 To measure the probabilistic eligibility, we fit the following logistic model: Pr( = 1) = (0 + 1 ) (2) where, is household ’s qualification statues (= 1 if eligible, 0 otherwise) under the full PMT for a specific MoLSA targeting rate (), is household’s PPMT score and () = /(1 − ). Figures A1 to A6 in Annex A present these probabilities for the national targeting rate of the bottom 18 percent ( = 18%) – the national poverty rate without adjusting for the SWIFT non-coverage of the most conflict affected districts. 11 As expected, a clear negative relationship between the probability of accurate referral and PPMT scores is observed, i.e., households that receive a low score using only the common set of variables (PPMT score) have a higher probability of having a low score according to the full PMT model and thus qualifying for MoLSA’s cash transfer program. For instance, households in the North with a PPMT score of 4.2 or less have a 90 percent probability of qualifying for the SSN program using the full MoLSA PMT to screen potential recipients (Table 5). Likewise, the referral threshold with 90 percent confidence of eligibility in Central-Southern region is 4.0 (Table 5). Given relatively low level of poverty in the Kurdistan region, almost no one from the region will qualify for the government program when targeting for the MoLSA CT is set at bottom 18 percent nationally (Table 5). As a consequence, there are no PPMT referral thresholds for the region at high confidence levels. 12 ), below which For practical purpose, these cutoff values can be interpreted as the referral threshold ( humanitarian caseloads would be referred to MoLSA for screening. For a given program size i.e. targeting bottom 18 percent nationally, changes with the level of referral confidence. However, there is a tradeoff between the confidence level and the number of referrals – for each program size number of household that would be referred to the government program decreased with confidence level. For example, if the PPMT score referral threshold is increased to 4.7 (VM-based score) or 4.8 (SEVAT-based score) in the North, it will entail screening more households but the referred households will have a 50 percent chance of being eligible for the government’s cash transfer program. For a given targeting rate, the PPMT referral thresholds corresponding to different confidence levels can be calculated for households in different regions of the country. Figures A1 to A6 in Annex A show the probability of referral to the MoLSA program (targeting the bottom 18 percent nationally) based on the VM- and SEVAT-based PPMT scores for the North, Central-South, and Kurdistan. These thresholds can be adjusted according to the capacity of the national SP system to screen referred households and absorb them into the system. We repeat the exercise for several other targeting rates namely: poorest 15 percent, poorest 25 percent, poorest 30 percent, and poorest 35 percent nationally. The PPMT cut-off World Bank (2019) for more. Hence, it is possible that, in addition to the smaller set, pseudo-PMT may be employing inferior proxies. 10 Refer to Box A 1, which presents a simple example for a quick understanding of how PPMT model and its referral probabilities can be calculated. 11 The continuous security issues prevented data collection from 14 qhadas (districts) and the final SWIFT sample covers only 106 of the 120 districts in the country. See Sharma & Wai-Poi (2019). 12 Currently there is no MoLSA’s CT program in Kurdistan region. However, a full PMT model was devised just like in the other three regions using the SWIFT 2017/18 data and same methodology for an emergency CT pilot that is currently being prepared by the Kurdistan Regional Government (KRG) MoLSA and the World Bank. 11 thresholds for these targeting rates are presented in Table A 2. Once the pseudo-PMT is implemented in the CLCI beneficiary database, one can use the information provided in this subsection to: (i) calculate the chances of being eligible for the MoLSA program for each beneficiary; and (ii) choose the optimal level of referral threshold, by referring to the graphs and tables presented in this subsection. Table 5: PPMT cut-off thresholds for referral to MoLSA program targeting the poorest 18 percent Referral confidence level = 90% 80% 70% 60% 50% North VM 4.2 4.5 4.6 4.7 4.7 SEVAT 4.3 4.4 4.6 4.6 4.8 Center-South VM 4.0 4.5 4.6 4.7 4.8 SEVAT 4.0 4.0 4.5 4.7 4.8 Kurdistan VM -- -- -- -- 4.6 SEVAT -- -- -- -- -- Source: Authors’ calculation using the Rapid Welfare Monitoring Survey (SWIFT) 2017-18. As with any targeting methodology, we also propose performance criteria for the PPMT models. Figure 7 illustrates the ways to calculate these referral statistics or the performance criteria for PPMT. While the performance of the PMT (inclusion and exclusion errors) is evaluated with respect to the “true” poverty status, the performance of the PPMT is evaluated with respect to the PMT, i.e., how well does PPMT reproduce the eligibility decisions of the PMT? In Figure 7, pseudo-PMT score, , is represented on the vertical axis, the full PMT score, , is on the horizontal axis. The vertical dotted line represents the full PMT targeting threshold, , which changes only when the overall targeting (e.g. poorest 18, 25, 30 percent nationally) is changed. The horizontal dotted line is a probabilistic and, in contrast, changes with the chosen confidence level of referral. Lower the threshold, higher is the referral accuracy, but higher may also be exclusion error as many who would have been eligible under the full PMT will not be referred. Cases in quadrant A in the figure are those that are eligible for the MoLSA program under the full PMT but would not be referred by the pseudo-PMT formula (e.g. this would be a failure to refer). Cases in quadrant B are correctly not referred (as they would not be eligible under the full PMT model); those in C are incorrectly referred (they are also not eligible under the full PMT model but the pseudo-PMT would indicate that they are); and those in D are successfully referred to the MoLSA program (both pseudo-PMT and full PMT indicate they are eligible). Based on these case types, one can calculate various referral statistics as shown in the figure. Statistic related to the “referral of ineligible cases” and the “non-referral of eligible cases” are comparable to the “inclusion error” and the “exclusion error” under the full-PMT. Since both the MoLSA PMT and pseudo-PMT are devised on the SWIFT data, one can estimate these statistics using the SWIFT data. However, given the differences between the survey data (nationally representative) and the database, these statistics would be more indicative and “theoretical” in nature. To assess the referral statistics in the humanitarian database, one would have to apply the PPMT and PMT on the database, perhaps by implementing a light survey allowing the implementation of PPMT and PMT formula on a small sample of humanitarian cases. 12 Figure 1: Performance criteria Eligibility threshold (Full PMT) Not referred to MoLSA Predicted consumption from pseudo-PMT A B Non-referral failure = A/(A+B) Referral threshold (Pseudo-PMT) Non-referral success = B/(A+B) Referral failure = C/(C+D) Referred to MoLSA Referral success = D/(C+D) D C Non-referral of eligible cases = A/(A+D) Referral of ineligible cases = C/(B+C) Predicted consumption from full PMT Absorbed by full PMT Rejected by full PMT Table 6 exhibits the performance/referral statistics for the North VM and SEVAT models for a MoLSA’s program targeting bottom 18 percent nationally. Minimizing the non-referral rate of eligible cases will prevent leaving cases that are in chronic poverty without state-distributed benefits after the humanitarian support stops. Nevertheless, the constraint is the national SP system’s ability to screen all the referred cases on time and take them into the system. Too many referrals may cause a backlog, while not all the referred cases will qualify for the MoLSA benefits. As seen in the table, the rate of non-referral of eligible cases falls dramatically as the referral confidence level is relaxed, while the referral rate of non-eligible cases rises but not by the same extent. There is a significant heterogeneity in the accuracy of the referral/non-referrals across the confidence levels. Decreasing confidence level implies increasing the referral threshold and thus referring more beneficiaries. Depending on the correlation between the pseudo-PMT and full-PMT scores, we may decrease the referral failure rate by lowering the confidence level or increasing threshold but may increase referral failure rates. Statistics for the Kurdistan and Center-South models are reported in Table A 3. Table 6: Performance criteria for VM and SEVAT north models for MoLSA program targeting the poorest 18 percent Referral confidence level = 90% 80% 70% 60% 50% VM Non-referral failure 23.1 21.0 20.5 17.3 15.9 Non-referral success 76.9 79.0 79.5 82.7 84.1 Referral failure 32.1 32.4 34.7 29.3 31.1 Referral success 67.9 67.6 65.3 70.7 68.9 Non-referral of eligible cases 91.8 79.5 76.2 61.5 54.5 Referral of non-eligible cases 1.2 3.2 4.1 5.2 6.6 13 SEVAT Non-referral failure 23.4 22.0 20.8 19.8 16.9 Non-referral success 76.6 78.0 79.2 80.2 83.1 Referral failure 45.8 36.2 35.7 31.2 28.3 Referral success 54.2 63.8 64.3 68.8 71.7 Non-referral of eligible cases 92.8 84.7 78.1 73.2 59.7 Referral of non-eligible cases 2.0 2.8 3.9 3.9 5.2 Source: Authors’ calculation using the Rapid Welfare Monitoring Survey (SWIFT) 2017-18. 4. Results This section presents the results from applying the pseudo-PMT models within the legacy database of current humanitarian cash assistance beneficiaries. The total revised legacy caseload at the time of the analysis was 37,457 households corresponding to 224,736 individuals. Initial analysis suggests strong potential for a sequenced referral process. CLCI’s CT program only operates in the districts affected by the conflict in the Northern governorates of Iraq. The CLCI’s legacy database and, hence, the results presented in this section are focused on the Northern region only. 13 As discussed under the methodology section, there are two dimensions of analytical variation. First, the likely size of the MoLSA budgets will determine how many new households can enter the program. Second, the degree of confidence of a successful referral is taken into account. Referring more people will increase the absolute number of households being determined eligible after MoLSA assessment but will also increase the absolute number of households being referred and not being determined eligible, potentially resulting in resentment and mismanaged expectations. Thus, thresholds were set for differing degrees of confidence that a given pseudo-PMT score would result in eligibility when the full MoLSA PMT is applied. To simplify the exposition, we present the results for targeting poorest 18 percent of the population under low (50 percent), medium (70 percent), and high (90 percent) referral confidence. The results for other targeting thresholds – 25 percent, 30 percent, and 35 percent – are presented in the Annex. 4.1 Likely Eligibility of Humanitarian Beneficiaries under MoLSA Targeting Initial aggregate results highlight the potential for both significant referral numbers and a sequenced strategy. Table 7 presents a first scenario with a smaller MoLSA budget for varying level of referral confidence. For a MoLSA program targeting the poorest 18 percent of the population and requiring a 90 percent confidence in the referral based on the pseudo-PMT, 24.3 percent of the existing humanitarian database would be eligible for referral. For the same level of targeting, 44.2 and 54.1 percent of the 13 At the time of the analysis, the new MCNA models were yet to be implemented for the humanitarian assistance eligibility. Besides it being limited to the Norther governorates, the beneficiary database was entirely based on the 2016 VM targeting methodology. Hence, the analysis in this section focuses on the 2016 VM pseudo-PMT model for the North only. Similar analytical exercise using the database resulting from employing the SEVAT targeting criteria and for other regions using other CWG partner’s database remains part of the next steps. 14 beneficiaries from the legacy database would be available for referral under the medium and low (70 and 50 percent) referral confidence. This will result in the smallest number of referrals – allowing time for MoLSA budgets and systems to absorb new beneficiaries over time – but also those households receiving humanitarian assistance who would be most likely to be eligible under MoLSA eligibility criteria. As the neediest cases are absorbed by the national system, the targeting threshold can be relaxed to 25, 30, or 35 percent of the poorest population. The percentage of cases from the humanitarian database that will be referred to MoLSA for screening under the relaxed targeting thresholds are shown in Annex Table A 1. In general, higher the targeting threshold and lower the referral confidence, the more cases will be referred. Table 7: Percentage of existing humanitarian database likely to be eligible for MoLSA program at different degrees of confidence and smaller program size Scenario Program Size Referral Confidence Percentage of Households High confidence, Poorest 18 percent 90 percent 24.3 percent medium-small program nationally Medium confidence, Poorest 18 percent 70 percent 44.2 percent medium-small program nationally Low confidence, medium- Poorest 18 percent 50 percent 54.1 percent small program nationally Data Source: CLCI Legacy Database. These results are important for two reasons. First, they suggest a large proportion of the humanitarian database could be potentially referred for MoLSA social protection programs. Second, a clear sequencing of smaller numbers with a high degree of likelihood to qualify could be referred first, with larger numbers being referred later when budgets expand. 4.2 Who Are the Likely Eligible? A Profile This sub-section profiles households from the CLCI legacy database that would be deemed eligible by the PPMT. The profile focuses on eligible households under the targeting of the bottom 18 percent nationally – the national poverty rate in 2017/18 before the adjustment for SWIFT non-coverage and likeliest coverage criteria. In particular, we analyze households that are eligible at 90 percent of confidence level, which we referred to as households with “High Likelihood of Eligibility (HLE)”. Analysis at 70 percent confidence, referred to as “Moderate Likelihood of Eligibility (MLE)” is discussed in Annex B. As reported in Table 7, households under HLE account for 24.3 percent of the legacy database, while 42.3 percent are eligible under the MLE scenario. Annex B also provides the profile of the households that are deemed ineligible even under the largest program and least confidence scenario explored in the paper i.e. poorest 35 percent nationally at 50 percent confidence level for comparison. Among the households with the highest likelihood of cross-model eligibility (90 percent confidence level; hereafter referred to as the HLE sample), a third are concentrated in Mosul, followed by other large urban centers such as Tikrit and Ramadi, followed by smaller urban centers and more rural areas affected by conflict such as Shirqat, Tel Afar, and Baiji. Not surprisingly, this is similar to the distribution in CLCI database that has greater coverage in these areas, which also have larger population sizes. 15 Thirty percent of the high likelihood sample is female headed, Table 8: Geographic Distribution of HLE which is roughly in line with CLCI vulnerability data broadly (not Households shown). The average household has 8 members (slightly higher District Distribution of than the national and CLCI average). Households Mosul 33% The average household size and member disaggregation does Tikrit 11% not change between female- and male-headed households, Ramadi 9% and nor by district – in the most populous districts the number Tel Afar 9% of women drops to 3 in Mosul and the number of children Shirqat 7% increases to 3 in Shirqat, Tel Afar, and Tikrit. The average head Kirkuk 5% of household age is 40 across the entire sample, dropping to 38 Qa’im 5% among female-headed households. In terms of head-of- Falluja 5% household specific vulnerabilities, 16 percent are disabled, and Baiji 5% 36 percent have a chronic illness, with female heads of Ana 4% household more often having chronic illnesses (59 percent). For Sinjar 3% Hamdaniya 2% 91 percent with a disability and 89 percent with a chronic All Others 2% illness, their condition prevents them from working. The Data Source: CLCI Legacy Database. average household has 1 disabled and 1 chronically ill member (not shown). Majority of the household heads have either no education, or have completed only primary education (Figure 8). This shifts markedly among female- headed households, among whom 50 percent received no education. Figure 2: Educational achievement of household head Figure 3: Household monthly income by number of working members. 60% IQD 400,000 50% IQD 350,000 40% IQD 300,000 30% IQD 250,000 20% IQD 200,000 10% 0% IQD 150,000 Undergraduate Secondary Primary Technical College None IQD 100,000 IQD 50,000 IQD 0 0 1 2 3 All Female-Headed Number of working members Data Source: CLCI Legacy Database. Data Source: CLCI Legacy Database. The average monthly income across the sample is IQD 137,041 / $115 (not shown). As expected, this increases with level of education and with household level of economic activity. Households with heads having an undergraduate degree earn 173% more than those having completed no education (not shown). 16 Households with 3 working members have income of IQD 250,000 / $211 more per month than households with no working members (Figure 9). Female-headed households, on average, have a much lower average monthly income, at IQD 85,631 / $72. By contrast, the mean income for male-headed households is IQD 156,781 / $132 (not shown). In general, households have expenditures that are greater than their incomes (Figure 10). Average total monthly expenditure for the sample is IQD 467,665 / $393, significantly greater than the average income of IQD 137,041, which translates to an average per capita monthly consumption of IQD 60,422 / $ 51. The gap between income and expenditure is partly explained by debt repayments (not included in the per capita consumption calculation), which do occur even among very low-income households, and the reliance on gifted in-kind food and non-food items and humanitarian aid to meet basic needs, in place of spending. The sample has relatively large debt – average of IQD 2,481,099 / $2,077. Much of the debt households accrue is to meet basic needs: more than four out of five (84 percent) resort to buying basic goods on credit as a coping strategy. Figure 4: Per capita income and consumption (IQD 0 50000 100000 150000 200000 250000 300000 350000 400000 450000 500000 Per Capita Income (IQD) Per Capita Consumption (IQD) Data Source: CLCI Legacy Database. Temporary unemployment among this sample is high: 72 percent of households reported no member of the household having worked in the previous 30 days. This ranges from 59 percent in Falluja to 93 percent in Tikrit. Among female-headed households, this increases to a mean of 86 percent. The number remained high over the period of assessment and changes inversely with per capita consumption (Figure 11). In terms of main sources of income, 31 percent of the households in the sample reported a temporary work, 22 percent state their community, friends, or family, and 16 percent reported loans. The remaining 24 17 percent state mosque donations, remittances, and government and social services payments as main source of income. Despite low incomes and low per capita consumption (which nevertheless exceeds incomes), food consumption scores (FCS) on average, across the sample, are acceptable. The mean is 40.5 (>35 is acceptable), dropping to a just-acceptable 36.1 among female-headed households with high standard deviation of 21.8 (Annex C). However, in certain districts households more often have borderline scores (25.5 – 35), for example in Ana, Tikrit and Qa’im (34.6, 30.2, and 26.7, respectively). One explanation for generally acceptable FCS might be the frequent incurrence of debt, whether as cash loans or goods on credit, which smooths food consumption. However, this needs additional data to verify. Most households in the sample (77 percent) resided in standard housing, with the remainder living in unfinished, abandoned, or public buildings, or informal tent settlements. Just under half (49 percent) share their housing with other families, with an average of 3 families per shared shelter. Over half (51 percent) of the sample have access to private hygiene and water facilities, with just under half (46 percent) saying they shared these with relatives; less than 2 percent said they had no access to water, sanitation, or hygiene facilities. Figure 5: Unemployment and per capita consumption (IQD) over the assessment period. 100% IQD 80,000 Unemployment (% of households with zero 90% IQD 70,000 80% IQD 60,000 Per capita consumption 70% working member) IQD 50,000 60% 50% IQD 40,000 40% IQD 30,000 30% IQD 20,000 20% IQD 10,000 10% 0% IQD 0 Jul-18 Jun-18 Nov-17 Nov-18 Apr-18 May-18 Dec-17 Aug-18 Dec-18 Jan-18 Feb-18 Mar-18 Sep-18 Jan-19 Sep-17 Oct-17 Oct-18 Unemployment Per Capita Consumption (IQD) Data Source: CLCI Legacy Database. 4.3 Likely Eligibility of Future Humanitarian Beneficiaries under MoLSA Targeting Following the completion of the MCNA survey in 2018, the Cash Working Group (CWG) updated the vulnerability assessment criteria for their cash transfer programs. Since the new models were yet to be fielded, there was no database similar to the legacy database to perform similar analysis. However, we 18 use the MCNA survey and calculate the SSN eligibility of the households in the survey that would be deem eligible under the new SEVAT model.14 Unlike the previous analysis of the actual database of the beneficiaries, this, using the survey, represents an exploratory analysis and the findings should be interpreted cautiously. Again, given the CLCI’s CT coverage and MCNA survey’s focus, we limit the cross- eligibility analysis to the sample from the Northern region that would be deem eligible under the new SEVAT targeting formula (about 14 percent). Table 9 presents the findings from this exploratory analysis. Results suggest much less overlap between MoLSA targeting and the new humanitarian targeting approach based on the MCNA survey data. Under a small program scenario i.e. targeting the poorest 18 percent nationally, the overlap varies between 0.7 to 13.2 percent depending on the level of confidence. Given equal targeting thresholds suggested by the PPMT, overlaps at high and medium confidence levels are same for a medium size program and increase to 15.4 percent at low confidence level. As expected, the overlap increase under a larger program scenario (between 3.3 to 19.9 percent) but remains significantly less than for the legacy database. Table 9. Percentage of likely CLCI beneficiaries in MCNA Survey that are likely to be eligible for MoLSA programs at different program sizes and degrees of confidence Scenario Program Size Referral Confidence Percentage of Households High confidence, small Poorest 18 percent 90 percent 0.7 percent program nationally Medium confidence, Poorest 18 percent 70 percent 6.5 percent small program nationally Low confidence, small Poorest 18 percent 50 percent 13.2 percent program nationally High confidence, medium Poorest 25 percent 90 percent 0.7 percent program nationally Medium confidence, Poorest 25 percent 70 percent 6.5 percent medium program nationally Low confidence, medium Poorest 25 percent 50 percent 15.4 percent program nationally High confidence, large Poorest 35 percent 90 percent 3.3 percent program nationally Medium confidence, Poorest 35 percent 70 percent 13.2 percent large program nationally Low confidence, large Poorest 35 percent 50 percent 19.9 percent program nationally Data Source: Multi Cluster Needs Assessment (MCNA) survey The lower level of overlap can potentially be explained by the fact that the likely eligibility exercise is carried out on a survey rather than the beneficiary’s database. First, even though there are a greater number of common variables in the two models, different enumeration may have led to questions being asked in different ways. For example, household size depends critically on what a respondent considers a household to be. For some, this may mean a family, for others multiple families living under the same 14 The updated SEVAT models provide three cut-off thresholds designed such that those deemed highly vulnerable would receive transfers up to three times per year, moderate vulnerable receive two times, while those deem just vulnerable receive a single transfer in a year. The overlap presented in Table 9 are for highly vulnerable households only (14 percent of the sample). Results are similar, decreases slightly, when including other two groups. 19 roof. When enumerators and manuals are different for different surveys, there can be systematic differences to the same question asked. Table 10 presents the statistical tests of the common variables between the two data sets (SWIFT and MCNA). All the common variables between the surveys are significantly different. The average household size in the SWIFT survey for the north is about 7, which is significantly greater (by one member) compared to the MCNA data. Similarly, there is significant difference in housing type, source of drinking water and household employment rate between the two surveys. As a result, the PPMT score in the MCNA data is slightly higher than in the SWIFT data. This implies that the PPMT would suggest an average person from the MCNA data is richer compared to an average person from the SWIFT data, which would directly impact the eligibility overlap using a common cut-off threshold. Table 10: Summary statistics of key variables by MCNA and SWIFT survey Survey Variables MCNA SWIFT Difference Home: house or flat 0.905 0.986 -0.082*** [0.294] [0.117] (0.012) Household size 5.946 6.993 -1.048*** [2.801] [3.274] (0.273) Employment rate (emp/hhsize) 0.180 0.209 -0.029*** [0.179] [0.144] (0.011) Water: public network tap 0.809 0.949 -0.141*** [0.393] [0.220] (0.023) Pseudo-PMT score (log) 5.176 5.118 0.058* [0.363] [0.374] (0.030) Observations 4012 300 Note: Standard deviations are in brackets, standard errors are in parentheses and significance levels are denoted as follows: *<0.10, **<0.05, ***<0.01 Second, the underlying populations of the SWIFT and MCNA surveys are quite different. SWIFT was designed to be nationally representative, albeit stratified to represent IDPs, who represent around 10 percent of the sample. MCNA was a vulnerability assessment and covers a significantly more vulnerable population. This may also explain some of the differences observed between the common variables in Table 10. These key differences in selection may have resulted in non-comparable samples which is affecting model comparison. 5. Conclusions and Next Steps With protracted displacement and delayed implementation of recovery and resilience activities, the humanitarian situation in Iraq is in transition. While 1.2 million Iraqis still remain displaced, more than 4.8 million of the 6.1 million that fled as a result of the ISIS conflict have returned to their areas of origin (IOM Iraq, 2021). Low prospect of employment/livelihood opportunities, lack of access to proper shelter/housing and other services, reduced social cohesion (community tensions and fears of population change) and security concerns (blocked return or perceived insecurity) form obstacles for safe and sustainable reintegration of some returnees (IOM Iraq, 2020a). While the Ministry of Migration and Displaced (MoMD) provides some support, access to social protection remains a challenge for both the displaced and returnees due to lack of civil documentation, inadequate assessment capacity of 20 authorities, and complicated registration procedures. The CWG’s MPCA programs have been instrumental in reaching the vulnerable households in areas most affected by the conflict and where the SSN coverage remains limited. Over a million conflict-affected households have been reached with the MPCA since the start of the conflict with ISIS. The short-term MPCA aid programs, however, will require a more permanent solution and eventual integration in some form with the government’s SSN programs to continue reaching the vulnerable households affected by the conflict. As an initial step, this paper proposes an analytical pseudo-PMT (PPMT) tool to bridge the targeting differences between the government SP program and the MPCA. The PPMT uses proxies that are common to both the humanitarian and the government CTP targeting formulas and provides each MCNA beneficiary’s probability of being eligible for the government’s CT program under different expansion scenarios. When applied to the existing humanitarian beneficiary database, the results suggest a strong source of referrals to the government’s SP system. Under a smaller program, targeting the poorest 15 or 18 percent nationally, between 24 to 54 percent of the existing humanitarian database would be available for referral to the government’s SSN programs with 90 to 50 percent (high and low) referral confidence levels. For a large program such as targeting the poorest 35 percent nationally, the referral rates increase to 40 to 73 percent for high and low referral confidence levels. Given a large proportion of the humanitarian database that could potentially be referred to MoLSA’s social protection programs, the results suggest a sequencing of smaller numbers with a high degree of likelihood to qualify could be referred first, with larger numbers being referred later when budgets expand. The MCNA households that are likely to be referred first, i.e., those with a high degree of likelihood (90 percent) under a smaller program expansion (targeting the poorest 18 percent nationally), are noticeable across several demographic and economic indicators. Majority of the early referrals are likely to come from urban centers such as Mosul, Tikrit and Ramadi with large concentration of IDPs and rural areas that experienced the conflict most - Shirqat, Tel Afar, and Baiji. Significant number of these households are headed by a female member (30 percent) and appear to have larger household sizes (8 members on average). They are also characterized by significant prevalence of disability and chronic illness among the head of the household. High unemployment, low income and expenditures but relatively high level of debt are other recognized characteristics among this group. 15 This initial analysis, however, has been based on probabilistic estimates from two different sources of data. Moreover, just because a household is likely to be eligible does not necessarily mean they are willing to be referred. Households with perceived ISIS affiliation may mean some households do not want to enter the formal government social protection system. To confirm the promising referral percentages from the desk analysis, but also determine the percentage of likely households who are also willing to be referred, a light field test is proposed. Under this test, the same households would be asked the full set of questions required to construct full MoLSA and VM/MCNA scores. This would provide stronger evidence of the apparent overlap between the two targeting systems. Moreover, willingness and perception questions could be asked to assess the combination of likely eligibility and willingness to be referred, which will ultimately determine how many potential MoLSA beneficiaries could come from the humanitarian database. 15 While the focus of the paper has been how close the PPMT scores are to the full PMT scores and how to overcome the difference, a more flexible ranking approach discussed in Annex D can be another alternative. 21 References Cash Consortium of Iraq. (2016). Beneficiary Vulnerability Analysis. Erbil: Cash Consortium of Iraq. IMF. (2021). Iraq Staff Report for the 2020 Article IV Consultation. Washington, D.C.: International Monetary Fund. IOM Iraq. (2020a). An Overview of Displacement in Iraq: DTM Integrated Location Assessment V, 2020. Baghdad: International Organization for Migration. IOM Iraq. (2020b). An Overview of Return Movements in Iraq: DTM Integrated Location Assessment V, 2020. Baghdad: International Organization for Migration. IOM Iraq. (2021). Displacement Tracking Matrix. International. Retrieved 7 21, 2021, from http://iraqdtm.iom.int/ Krishnan, N., & Olivieri, S. D. (2016). Losing the gains of the past : The welfare and distributional impacts of the twin crises in Iraq 2014. Policy Research working paper, no. WPS 7567. Sharma, D., & Wai-Poi, M. G. (2019). Arrested Development : Conflict, Displacement, and Welfare in Iraq : Arrested Development - Conflict Displacement and Welfare in Iraq. World Bank. Washington, D.C.: World Bank Group. World Bank. (2014). The Unfulfilled Promise of Oil and Growth: Poverty, Inclusion and Welfare in Iraq 2007-2012. Washington DC: World Bank. World Bank. (2019). Updating the Proxy Means Test Formula for Iraq – A Technical Note. Washington D.C: World Bank. 22 Annex A: Additional Results Figure A. 1: Probability of referral to MoLSA program Figure A. 2: Probability of referral to MoLSA program which targets the poorest 18 percent based on the VM targeting the poorest 18 percent based on the new pseudo-PMT for the Northern region. SEVAT pseudo-PMT for the Northern region. 1 1 .8 .8 .6 .6 Probability Probability .4 .4 .2 .2 0 0 3.5 4 4.5 5 5.5 6 6.5 4 4.5 5 5.5 6 6.5 Predicted log expenditure from pseudo-model Predicted log expenditure from pseudo-model Note: Probabilities are calculated running logistic regression of poverty status under full-PMT on predicted expenditure using the respective pseudo-PMT models. Data Source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. Figure A. 3: Probability of referral to MoLSA program Figure A. 4: Probability of referral to MoLSA program targeting the poorest 18 percent based on the VM targeting the poorest 18 percent based on the SEVAT pseudo-PMT for the Center-Southern region. pseudo-PMT for the Central-Southern region. 1 1 .8 .8 .6 .6 Probability Probability .4 .4 .2 .2 0 0 3.5 4 4.5 5 5.5 6 6.5 3.5 4 4.5 5 5.5 6 6.5 Predicted log expenditure from pseudo-model Predicted log expenditure from pseudo-model Note: Probabilities are calculated running logistic regression of poverty status under full-PMT on predicted expenditure using the respective pseudo-PMT models. Data Source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. 23 Figure A. 5: Probability of referral to MoLSA program Figure A. 6: Probability of referral to MoLSA program targeting the poorest 18 percent based on the VM targeting the poorest 18 percent based on the SEVAT pseudo-PMT for Kurdistan region. pseudo-PMT for Kurdistan region. 1 1 .8 .8 .6 .6 Probability Probability .4 .4 .2 .2 0 0 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 4.6 4.8 5 5.2 5.4 5.6 5.8 6 6.2 Predicted log expenditure from pseudo-model Predicted log expenditure from pseudo-model Note: Probabilities are calculated running logistic regression of poverty status under full-PMT on predicted expenditure using the respective pseudo-PMT models. Data Source: Rapid Welfare Monitoring Survey (SWIFT) 2017-18. Box A 1: A simplified example of PPMT procedure. This box presents a simplified example of how PPMT creation and referrals are determined. In the example, the national PMT model uses four proxies, while the humanitarian PMT has three. The PPMT model then would be created using the two common variables between the models, i.e., years of education and access to good water. Once the PPMT weights are calculated using the household survey data that was used to devise the national PMT and running an ordinary least squares (OLS) regression of household consumption on the two common indicators, we can calculate households’ eligibility using the PPMT. As presented in the scenario below, assume the population of three households and all three have equal PPMT scores of 1.0 but have different full PMT scores. Also, assume the eligibility threshold of 1.05 for the full PMT, below which households would be considered monetary poor and eligible for the government cash transfer program. Under such eligibility criteria, only two households, the first and third with the national PMT scores of 0.97 and 1.02, would qualify for the national program. This information 24 is sufficient to calculate the probability of referral given one’s PPMT score. In the example, households with PPMT scores of 1.0 would have 66.7 percent chance of being eligible under the full national PMT. Using a nationally representative household survey, one would be able to calculate the probabilities for numerous other PPMT scores. Table A 1: Percentage of existing humanitarian database likely to be eligible for MoLSA program at different degrees of confidence and program size. Scenario Program Size Referral Confidence Percentage of Households High confidence, small Poorest 15 percent 90 percent 24.3 percent program nationally Medium confidence, Poorest 15 percent 70 percent 40.4 percent small program nationally Low confidence, small Poorest 15 percent 50 percent 54.1 percent program nationally High confidence, medium Poorest 25 percent 90 percent 34.0 percent program nationally Medium confidence, Poorest 25 percent 70 percent 54.1 percent medium program nationally Low confidence, medium Poorest 25 percent 50 percent 68.6 percent program nationally High confidence, Poorest 30 percent 90 percent 34.0 percent medium-large program nationally Medium confidence, Poorest 30 percent 70 percent 54.1 percent medium-large program nationally Low confidence, medium- Poorest 30 percent 50 percent 73.1 percent large program nationally High confidence, large Poorest 35 percent 90 percent 40.4 percent program nationally Medium confidence, Poorest 35 percent 70 percent 50.8 percent large program nationally Low confidence, large Poorest 35 percent 50 percent 73.3 percent program nationally Data Source: CLCI Legacy Database. Table A 2: PPMT cut-off thresholds for referral to MoLSA program at different targeting. Referral confidence level = 90% 80% 70% 60% 50% Bottom 15% North VM 4.2 4.4 4.5 4.6 4.7 SEVAT 4.2 4.4 4.5 4.6 4.7 Center-South VM 4.0 4.0 4.5 4.6 4.7 SEVAT 4.0 4.0 4.5 4.6 4.7 Kurdistan VM -- -- -- -- -- SEVAT -- -- -- -- -- Bottom 25% 25 North VM 4.4 4.6 4.7 4.8 4.9 SEVAT 4.3 4.5 4.6 4.8 4.9 Center-South VM 4.0 4.5 4.6 4.8 4.9 SEVAT 4.0 4.0 4.6 4.7 4.8 Kurdistan VM -- -- 4.6 4.7 4.7 SEVAT -- -- -- 4.6 4.7 Bottom 30% North VM 4.4 4.6 4.7 4.8 5.0 SEVAT 4.3 4.5 4.7 4.8 4.9 Center-South VM 4.0 4.6 4.7 4.8 4.9 SEVAT 4.0 4.5 4.7 4.8 4.9 Kurdistan VM -- 4.6 4.7 4.7 4.9 SEVAT -- -- 4.6 4.7 4.7 Bottom 35% North VM 4.5 4.7 4.9 5.0 5.0 SEVAT 4.5 4.7 4.8 4.9 5.0 Center-South VM 4.6 4.7 4.8 4.9 5.0 SEVAT 4.0 4.6 4.8 4.9 5.0 Kurdistan VM 4.6 4.7 4.7 4.9 5.0 SEVAT -- 4.6 4.7 4.9 5.0 Table A 3: Performance criteria for VM and SEVAT models for MoLSA program targeting the poorest 18 percent nationally. Center-South VM Non-referral failure 18.5 18.4 17.9 17.2 16.4 Non-referral success 81.5 81.6 82.1 82.8 83.6 Referral failure 0.0 25.5 38.4 32.9 40.3 Referral success 100.0 74.5 61.6 67.1 59.7 Failure to refer 96.4 95.6 91.9 87.5 81.1 Failure to non-refer 0.0 0.4 1.2 1.4 3.0 SEVAT Non-referral failure 18.5 18.5 18.3 17.4 16.0 Non-referral success 81.5 81.5 81.7 82.6 84.0 Referral failure 0.0 0.0 22.6 32.8 29.1 26 Referral success 100.0 100.0 77.4 67.2 70.9 Failure to refer 96.4 96.4 94.9 88.5 79.8 Failure to non-refer 0.0 0.0 0.4 1.3 1.9 Kurdistan VM Non-referral failure 5.5 5.5 5.5 5.5 5.1 Non-referral success 94.5 94.5 94.5 94.5 94.9 Referral failure -- -- -- -- 0.0 Referral success -- -- -- -- 100.0 Failure to refer 100.0 100.0 100.0 100.0 92.7 Failure to non-refer 0.0 0.0 0.0 0.0 0.0 SEVAT Non-referral failure 5.5 5.5 5.5 5.5 5.5 Non-referral success 94.5 94.5 94.5 94.5 94.5 Referral failure -- -- -- -- -- Referral success -- -- -- -- -- Failure to refer 100.0 100.0 100.0 100.0 100.0 Failure to non-refer 0.0 0.0 0.0 0.0 0.0 27 Table A 4: Selected characteristics (HLE) Female-headed Household FCS Score Per Capita Consumption (Monthly) Mean 36.12896574 Mean 60421.83872 Standard Error 0.387551724 Standard Error 567.3301235 Median 34.5 Median 46920 Mode 0 Mode 0 Standard Deviation 21.75819062 Standard Deviation 54146.60044 Range 112 Range 1167000.4 Minimum 0 Minimum 0 Maximum 112 Maximum 1167000.4 Sum 113878.5 Sum 550382528.9 Count 3152 Count 9109 Confidence Level (95.0%) 0.759879304 Confidence Level (95.0%) 1112.094396 Total Household Debt Household Size Mean 2481099.479 Mean 7.6532 Standard Error 89861.70624 Standard Error 0.026385 Median 1000000 Median 7 Mode 1000000 Mode 7 Standard Deviation 8576498.411 Standard Deviation 2.518257 Range 375000000 Range 24 Minimum 0 Minimum 4 Maximum 375000000 Maximum 28 Sum 22600335155 Sum 69713 Count 9109 Count 9109 Confidence Level (95.0%) 176149.1163 Confidence Level (95.0%) 0.051721 Coping Strategy Index Score Household Income (Monthly) Mean 18.02491 Mean 137056.3374 Standard Error 0.084621 Standard Error 2186.947803 Median 17.7 Median 75000 Mode 35.77 Mode 0 Standard Deviation 8.076278 Standard Deviation 208690.2891 Range 35.77 Range 2000000 Minimum 0 Minimum 0 Maximum 35.77 Maximum 2000000 Sum 164189 Sum 1248035008 Count 9109 Count 9106 Confidence Level (95.0%) 0.165875 Confidence Level (95.0%) 4286.908805 28 Annex B: Profile of eligible households under a Moderate Likelihood of Eligibility (MLE) and non-eligible households even under the most generous program scenario Moderate Likelihood of Eligibility (70 % confidence level) The sample with moderate likelihood of cross-model eligibility (18 percent national targeting and at 70 percent confidence; hereafter referred to the MLE sample) is across numerous indicators similar to the high likelihood sample. The geographic distribution of households is almost identical, again factoring in the bias in the data towards CLCI areas of intervention (Table B. 1). Table B. 1: Geographic distribution As with the higher likelihood sample, 30 percent are female of households headed. The average household has 7 members, 2 women, and District Distribution of 5 children per household, including 2 non-family children. The Households average head of household age is again 40, dropping to 39 Mosul 35% among female heads of household. The rates of heads of Tikrit 11% households with disabilities and chronic illnesses is almost the Ramadi 9% same: 16 percent and 35 percent respectively. As shown in Tel Afar 8% Table B. 2, levels of education are likewise identical, as is the Shirqat 8% disparity between the average levels of schooling and those Kirkuk 6% among female heads of household. Qa’im 5% Baiji 5% Average monthly incomes among this sample were slightly Fallujah 4% higher, at IQD 147,590 / $124, as is average monthly Ana 4% expenditure at IQD 477,198 / $400, and average per capita consumption at IQD 65,243 / $55. The same gap persists among Sinjar 2% this sample between monthly per capita income and Hamdaniya 2% expenditure, as do the high levels of household debt, at IQD All Others 2% Table B. 2: Head of Household Level of 2,537,805 / $2,132. Those with undergraduate degrees earn Education (MLE) considerably more than those with no education, as do those with at least 1 working household member (Figure B. 1). Schooling All Female Temporary unemployment among this sample is lower, with 66 percent saying no member of the household had worked in the None 34% 50% past 30 days. The primary source of income increases slightly Primary 48% 45% towards temporary work (35 percent), with slightly fewer (20 Secondary 12% 4% percent) relying on community, friends, or family. Undergraduate 5% 1% Food consumption scores are on average slightly higher and Technical College 1% 0% remain acceptable, with a mean of 41.5, dropping to 36.6 among female-headed households. Most (80 percent) similarly report living in standard housing, with just over half (52 percent) reporting sharing their shelter with an average of 3 families. And as with the high likelihood sample, 97 percent report having access to private water, sanitation, and hygiene facilities, or sharing these with relatives. 29 Figure B. 1: Average income, education, and work IQD 350,000 IQD 300,000 IQD 250,000 IQD 200,000 IQD 150,000 IQD 100,000 IQD 50,000 IQD 0 No Education Undergraduate Education No Working Members One Working Member Non-Eligible Households (outside the PPMT model) Among the sample who are deemed ineligible even under the largest and least confidence level (35 percent targeting and 50 percent confidence), the differences across numerous demographic and economic indicators become more noticeable. Geographically, the distribution remains similar, with an increase of households in Baiji and a marked decrease in Tikrit. There are slightly fewer female-headed households, at 28 percent. The average household in this sample is smaller with 4 members, with lower averages of 2 women and 2 children. Average monthly incomes are higher, at IQD 178,789 / $150, but monthly expenditure is on average lower, at IQD 446,999 / $375. Average incomes among female-headed households are lower than the average, at IQD 120,272 / $101. Income among male-headed households was IQD 187,572 / $158. Levels of debt are also on average slightly lower, at IQD 2,362,980 / $1,985. The smaller average household size appears to have a positive impact on per capita consumption. In this sample, average per capita consumption is IQD 110,981 / $93, compared to IQD 60,422 / $51 in the HLE sample. Median per capita consumption in this sample is IQD 82,714 / $69, compared with IQD 46,920 / $39 in the HLE sample. Per capita consumption is smoother across the distribution compared with the other samples (Figure B. 2). However, the non-eligible households have only 5 percent less debt than the HLE sample indicates these households are still reliant on borrowing to meet basic needs. Households within this sample more often cite temporary work as their primary source of income (38 percent; 31 percent in HLE sample) and less often cite their community, friends, or family (18 percent; 22 percent in HLE sample). 30 Employment among this sample was higher, with 60 percent citing no member of the household had worked in the past 30 days, compared with 72 percent in the HLE sample, and with positive trends towards greater temporary and regular work compared with the HLE and MLE samples (Figure B. 3). Lower unemployment held true across districts too, ranging from 54 percent in Mosul to 76 percent in Fallujah. Female-headed households, despite still reporting lower economic activity than the average, fared better than the HLE sample, with 66 percent reporting no working household member. Figure B. 2: Per capita expenditure (IQD) 0 50000 100000 150000 200000 250000 300000 350000 400000 450000 500000 Per Capita Consumption (IQD) 90% CL Per Capita Consumption (IQD) 70% CL Per Capita Consumption (IQD) Ineligible Figure B. 3: Type of employment of household members Non-Eligible Moderate Likelihood High Likelihood 0% 10% 20% 30% 40% 50% 60% 70% 80% Regular Work Temporary Work No Work 31 Education level of household head did not differ from the HLE or MLE samples significantly, with slightly less primary school completion, slightly higher rates of no schooling, but slightly higher rates of secondary and undergraduate completion. In this sample, the disparity between the average and female levels of education is greater than in the HLE and MLE samples, as shown in Figure B. 4 below. Figure B. 4: Education of household head 70% 60% 50% 40% 30% 20% 10% 0% None Primary Secondary Undergraduate Technical College All Female-Headed Finally, in the non-eligible sample the average FCS was higher than for the HLE and MLE samples, at 43.7 across the sample and 39.6 among female-headed households. The better represented districts in the sample that had borderline or poor FCS were Fallujah, Qa’im, and Tikrit. Annex C: Social Protection and Humanitarian Assistance in Iraq National Social Protection System Since 2006 the Government of Iraq (GOI) has initiated poverty reduction efforts, which culminated in the launch of the first Poverty Reduction Strategy (PRS 1) by the Ministry of Planning (MOP) in 2010. PRS 1 refocused social welfare expenditure away from massive government subsidies for consumption to an expanded investment in human productivity and targeted transfers to the poor. Given the urgency of new challenges from the economic and political developments in Iraq, namely falling oil prices and the Islamic State of Iraq and the Levant (ISIL) occupation and liberation, MOP has recently launched the Second PRS (2018-2022), which focuses on creating opportunities for sustainable income; empowerment and building human capital; and establishing effective social safety nets (SSN). The new strategy pays particular attention to the post-conflict emergency needs, including recovery of Internally Displaced Persons (IDPs) and returnees, as well as physical rehabilitation. Along with these poverty reduction efforts, the GOI has embarked on a comprehensive social protection reform that introduced significant improvements to the existing system by promoting equity, resilience, 32 and opportunities for the Iraqi people. The 2015-2019 Social Protection Framework guides the Ministry of Labor and Social Affairs’ (MOLSA) reforms in the three pillars of social safety nets, social insurance, and labor markets. With the support of the World Bank, a major achievement was the issuance of the new Social Protection Law of Iraq (11/2014), which provided for the shift from categorical to poverty targeting in social assistance, hence improving outreach to the poor. The Social Protection Network (SPN) now reaches 1.36 million of the poorest households with monthly electronic cash transfers. In addition, MOLSA, in collaboration with UNICEF and the World Bank, launched a pilot Incentivized Conditional Cash Transfer (ICCT) in one of the poorest neighborhoods of Baghdad. The pilot ICCT, which focuses on a small number of poor households who are already benefiting from the SPN, is undergoing an impact evaluation that will shed light on whether it leads to improved human capital for poor families through increased access to education and health services for school-age children, children under-five, and women of child-bearing age. MOLSA plans to expand the ICCT to additional governorates through the World Bank supported Emergency Social Stability and Resilience Project (ESSRP). On the social insurance front, the World Bank has supported the GoI to introduce a new social insurance law, currently under discussion in Parliament. The new law integrates the public and private sector pension schemes and is hoped to have positive implications on labor mobility and fiscal rationalization of the pension fund. Finally, MOLSA has recently started exploring possible interventions that aim to address the issue of unemployment in Iraq, especially for youth and women. Beyond the social protection efforts led by MOLSA, the Public Distribution System (PDS) under the Ministry of Trade (MOT) is the largest SSN in Iraq, covering 95 percent of the population. In fact, PDS is the largest source of calories for the poor. WFP has taken a lead role in supporting MOT to modernize the PDS database, as well as to implement biometrics to reform the PDS Ration Cards system. In addition, WFP is currently piloting a project that will ensure interoperability of the PDS database and that of other social protection beneficiaries implemented by other relevant ministries. It is hoped that this pilot project leads to improved management of PDS and increased efficiency of service delivery to Iraqi citizens. Humanitarian Assistance The humanitarian situation in Iraq is in a transition phase, evolving from the response to the conflict with ISIL to a humanitarian context with unforeseen protracted displacement and delayed implementation of recovery and resilience activities. While 68 percent of the 6 million IDPs have returned to their areas of origin, those who were among the initial wave of returnees have in some cases found destroyed or damaged houses, unexploded ordinances, absence of livelihoods and services, lack of social cohesion (community tensions and fears of population change) and security concerns (blocked return or perceived insecurity). Of the approximately 1.9 million who remain displaced from their area of origin – over half for more than three years – approximately 65 percent of those based both in and out of camps have indicated that they have no intention to return to their areas of origin in the next 12 months. A reported 12 percent of all IDPs (approximately 225,000) have expressed a desire to locally integrate into their pace of displacement, especially if there are family connections or wider availability of jobs and services. It is becoming clear that a significant majority of remaining IDPs may not return to their area of origin and data collection may be underestimating the number of those who wish to integrate locally or resettle in the long-term. While the Ministry of Migration and Displaced (MOMD) is providing support to affected populations, access to relevant social protection networks remains challenging for both displaced people and returnees. Obstacles include a lack of necessary civil documentation for enrollment in social programs, inadequate assessment capacity of the authorities and complicated registration procedures. A responsible 33 transition is needed towards government and international development planning frameworks, the RRP, FFS and UNDAF. As a result of the humanitarian crisis in Iraq, the poverty rate among IDPs is estimated at 38 percent, twice as high as for the rest of the population. 16 Moreover, the costs of the war are daunting. A Damage and Needs Assessment (DNA) which was conducted in 2017 revealed damages worth $45.7 billion and needs amounting to $88.2 billion. 17 In this context, the Iraqi conflict has prompted a large-scale development and humanitarian response to meet the needs of IDPs and returnees. The Kuwait International Conference for Reconstruction of Iraq held in January 2018 raised over $30 billion in pledges for reconstruction, recovery and development needs. At the same time, humanitarian agencies and organizations looking at interventions related to social protection include UNICEF, WFP, UNHCR, the Cash Consortium for Iraq (CLCI), and REACH, along with donors (DFID, ECHO, OFDA/FFP) and OCHA. For example, UNICEF Iraq has been working with the GOI, civil society partners and the private sector to deliver unconditional education-focused humanitarian cash assistance in the governorates of Dohuk, Erbil, Anbar and Ninawa. The objective of the intervention is to remove financial and structural barriers to school enrollment and retention among vulnerable IDP, refugee and returnee children using a vulnerability lens. The project has supported over 15,000 children over the last two years. The Cash Consortium for Iraq was formed in March 2015 by the Danish Refugee Council (DRC), the International Rescue Committee (IRC), the Norwegian Refugee Council (NRC) and Mercy Corps as lead agency. The CLCI was formalized with the aims of enhancing the impact of multi-purpose cash assistance (MPCA) by building a harmonized approach to MPCA delivery, fostering closer operational coordination, and expanding geographic reach. Oxfam joined the CLCI in 2016 as the fifth partner. The CLCI has received funding for MPCA from Global Affairs Canada (GAC), ECHO, the United States’ Office of Foreign Disaster Assistance (OFDA) and Food For Peace (FFP), the United Kingdom’s Department for International Development (DFID), the United Nations Office for the Coordination of Humanitarian Affairs (UNOCHA) and the Iraq Humanitarian Pooled Fund (IHPF), with a current funding portfolio of over $40,000,000 USD. The CLCI has reached over 75,000 households (approximately 450,000 individuals) with MPCA since 2015 and is currently operating in the five largest conflict-affected governorates in Iraq. Cash Transfer Programs have been used in Iraq since 2014, initially for Syrian refugees living in camp and out-of-camp settings across northern Iraq and the KRI. In late-2014 and 2015, the internal displacement following the conflict with ISIS led to cash and voucher programs being implemented at a larger scale by NGOs, UNHCR, and WFP. The Iraq Cash Working Group (CWG) was formed in 2014 as a ‘semi-cluster’ to coordinate the growing MPCA response and was also given a seat at the Inter-Cluster Coordination Group (ICCG) and, following advocacy by MPCA actors and the CLCI Steering Committee, developing a standalone MPCA chapter in the 2015 Humanitarian Response Plan (HRP). MPCA has had a chapter in the Iraq HRP every year since. Over a million conflict-affected households have been reached with MPCA since the start of the conflict with ISIS. The tools developed by the CLCI for MPCA vulnerability assessment, scoring, and Post Distribution Monitoring have been endorsed by the CWG and are widely used by other MPCA actors including I/NGOs as well as UN agencies. In April 2018, early conversations on a responsible transition to the SPN began to build momentum through a two-day Social Protection workshop convened by DFID, with MoLSA, the World Bank, the CLCI, 16 See World Bank (forthcoming) Arrested Development: Conflict, Displacement and Welfare in Iraq. 17 See World Bank (2017) Disaster Needs Assessment. 34 the CWG and key UN agencies in attendance. It was recognized that, among Iraq’s myriad SPN services, the MoLSA SSN was the best placed institution to meet the needs of conflict-affected households vulnerable to future shocks or poverty, through its targeted cash transfers. The first outcome from the workshop was an Action Plan that identified key areas in need of stakeholder focus, including targeting, referrals, and information management. During the workshop, the World Bank presented the PMT targeting models used by MoLSA and the CLCI presented the PMT utilized by the CLCI and other members of the CWG. The two eligibility models were shown to have a significant degree of overlap, and this overlap in criteria suggested that targeting is a good basis from which to seek closer alignment and establish ways of working during the transition. Annex D: Ranking Approach Besides the hard cut-off thresholds, the PPMT scores, an estimate of household well-being, directly can be exploited to assess the cross-eligibility. While the focus of the paper has been how close the PPMT scores are to the full PMT scores and how to overcome the difference using the probabilistic approach, one can focus on how well the PPMT ranks households compared to the full PMT. If the PPMT sufficiently replicates households’ ordering/ranking under the MoLSA’s full PMT, the PPMT-scores can be used as the running well-being variable. For instance, if we knew that for a given budget 25 percent of the households form a group/region would be eligible using the full PMT, we can simply refer the 25 percent with the lowest PPMT scores from the group/region to the MoLSA’s program. This more flexible approach is likely to overcome the differences that will arise from the disparities in data set that was used to devise the PPMT and the one where it is applied to. Differences in survey methodologies may mean that a same variable between two surveys may not be comparable. Figure D.6: Scatter plot of MoLSA full PMT and Pseudo- Figure D.7: Scatter plot of MoLSA full PMT and Pseudo- PMT VM North PMT SEVAT North Spearman's Rank Correlation Coefficients = 0.763* Spearman's Rank Correlation Coefficients = 0.688* 6.5 6.5 6 6 Score from pseudo-model Score from pseudo-model 5.5 5.5 5 5 4.5 4.5 4 4 3.5 4 4.5 5 5.5 6 6.5 4 4.5 5 5.5 6 6.5 Predicted log expenditure from full MoLSA model Predicted log expenditure from full MoLSA model Source: Authors’ calculation using the Rapid Welfare Monitoring Survey (SWIFT) 2017-18. * Indicates statistical significance at the 1% level. The underlaying assumption for such strategy is that the PPMT orders households similarly as the full PMT. While a light survey where a same households would provide information required to construct the full MoLSA and VM/SEVAT scores would be ideal to test the assumption, we can theoretically assess the 35 ordering performance of the PPMT in the SWIFT data. Figures D.1 to D.6 exhibit the rank order correlation coefficients and scatter plots of PPMT and PMT scores. The strength of the correlation is relatively strong across all models. The rank correlation between the PMT and PPMT is strongest in the North (0.763 for VM and 0.688 for SEVAT), followed by the models for Kurdistan region (0.683 for VM and 0.619 for SEVAT) and the Center-South models (0.588 for VM and 0.562 for SEVAT). The strong correlation coefficient and visual from the scatter plots suggest, at least theoretically, the PPMT replicates the ordering of household from the full PMT relatively well. Figure D.8: Scatter plot of MoLSA full PMT and Pseudo- Figure D.9: Scatter plot of MoLSA full PMT and Pseudo- PMT VM Center-South PMT SEVAT Center-South Spearman's Rank Correlation Coefficients = 0.588* Spearman's Rank Correlation Coefficients = 0.562* 6.5 6.5 6 6 Score from pseudo-model Score from pseudo-model 5.5 5.5 5 5 4.5 4.5 4 4 3.5 3.5 4 4.5 5 5.5 6 6.5 4 4.5 5 5.5 6 6.5 Predicted log expenditure from full MoLSA model Predicted log expenditure from full MoLSA model Source: Authors’ calculation using the Rapid Welfare Monitoring Survey (SWIFT) 2017-18. * Indicates statistical significance at the 1% level. Figure D.10: Scatter plot of MoLSA full PMT and Figure D.11: Scatter plot of MoLSA full PMT and Pseudo-PMT VM Kurdistan Pseudo-PMT SEVAT Kurdistan Spearman's Rank Correlation Coefficients = 0.683* Spearman's Rank Correlation Coefficients = 0.619* 6.2 6.2 6 6 Score from pseudo-model Score from pseudo-model 5 5.2 5.4 5.6 5.8 5 5.2 5.4 5.6 5.8 4.6 4.8 4.6 4.8 4 4.5 5 5.5 6 6.5 7 4 4.5 5 5.5 6 6.5 7 Predicted log expenditure from full MoLSA model Predicted log expenditure from full MoLSA model Source: Authors’ calculation using the Rapid Welfare Monitoring Survey (SWIFT) 2017-18. * Indicates statistical significance at the 1% level. 36