Policy Research Working Paper 9465 In Light of What They Know How Do Local Leaders Make Targeting Decisions? Ervin Dervisevic Seth Garz Aneesh Mannava Elizaveta Perova East Asia and the Pacific Region Gender Innovation Lab & Office of the Chief Economist November 2020 Policy Research Working Paper 9465 Abstract This paper analyzes how local leaders make targeting deci- shocks—indicators that are likely to differ in their observ- sions in the context of a public workfare program in the ability to village heads and could plausibly be associated Lao People’s Democratic Republic. The study finds that vil- with need for public support. The study finds that village lage heads are progressive in their targeting, prioritizing the heads’ perceptions, as revealed through the ranking exercise, poorer households in their villages. The study benchmarks differ substantially from actual levels reported in surveys of this decentralized selection to the common alternative the villagers themselves. The study then uses a data-driven proxy means test method and finds that village heads are at machine learning approach to identify the predictors of least as progressive as a proxy means test method approach. village head selection. It concludes that village heads rely To illuminate what poverty-related information village on a combination of easily observable household character- heads could plausibly be incorporating into their internal istics, forming a holistic impression of household welfare, selection decisions, the study designs and administers a rather than specific indicators like actual land ownership, set of exercises for village heads to rank villagers on land nutrition, or economic shocks. ownership, access to nutrition, and experience with recent This paper is a product of the Gender Innovation Lab and the Office of the Chief Economist, East Asia and the Pacific Region. 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 eperova@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 In Light of What They Know: How Do Local Leaders Make Targeting Decisions? 1 Ervin Dervisevic*, Seth Garz^, Aneesh Mannava* and Elizaveta Perova*~ JEL classification: I38, D73, O12 Keywords: Targeting Social protection Public works Decentralization Local leaders Measurement 1 We gratefully acknowledge funding from the Umbrella Facility for Gender Equality (UFGE) and Australia’s DFAT to carry out this work. The data collection protocols were approved by Health Media Labs IRB. We thank Alana Teman for excellent field coordination, and Darathip Sounthavong, Vanvilay Phommalath and Khanithalath Koummarasy for the management of data collection process. We are grateful to Hatthachan, Hongngeun Sadettan and other members of Laos PRF for providing detailed information about the RMG program and sharing their insights about our findings. We thank Aline Coudouel, Patrick Premand, Stephanie Kuttner, Erik Johnson, Norman Loayza, Sybounheung Phandanouvong, Kathleen Beegle, Aaditya Mattoo and East Asia and Pacific Chief Economist Office seminar participants for helpful comments. All errors are our own. The IE was carried out by the East Asia and Pacific Chief Economist Office, under PASA activity P156793. Elizaveta Perova was the TTL of the PASA. Erik Caldwell Johnson is the TTL of the Poverty Reduction Fund III AF (P168620) operation, which supports the Laos Road Maintenance Groups project. * World Bank; ^ Bill and Melinda Gates Foundation; ~ corresponding author: eperova@worldbank.org 1. Introduction Targeting of social programs by local leaders, fully or in combination with some other targeting method, has been at the core of development policies and programs for many years. The discussion of the merits and perils of this approach to targeting has been comparably long. Recently, the COVID-19 pandemic has necessitated quick and effective emergency assistance delivery to those most affected, bringing new relevance to the old debate. Indeed, there are several strong arguments in favor of this approach. Local leaders – individuals deeply embedded in the context of the lives of potential beneficiaries – may have better information about the community and relative needs of its residents (Conning and Kevane 2002; Subbarao et al. 2012). Reliance on local leaders also presents an opportunity to save resources needed for collecting information to implement alternative targeting methods, such as proxy means tests, which are often more data intensive (Alix-Garcia et al., 2019). Finally, selection of beneficiaries by a trusted member of the community may be easier to understand than selection based on a more technical approach, 2 and thus result in greater acceptance of targeting results by the community (Alatas et al., 2012). On the other hand, vesting power of beneficiary selection in the hands of several community members may result in nepotism and elite capture. This has been an important concern for reliance on targeting by local leaders. There is no consensus in the literature on the extent of and welfare losses from elite capture in targeting of social programs by local leaders. Kilic et al. (2013) find evidence of nepotism when analyzing the distributional effectiveness of the Malawi Farm Input Subsidy Programme (FISP). Their results suggest that the relatively well-off and locally well-connected are more likely to access the program where targeting decisions are largely made by village chiefs. Analyzing the same program but using a different measure of poverty, Basurto, Dupas and Robinson (2019) suggest that the consequences of nepotism are likely to be small: a PMT would not do better. This result corroborates an earlier finding from Alatas et al. (2013) in Indonesia, who suggest that although elite capture exists in the distribution of village welfare programs, eliminating it entirely would only marginally improve welfare gains from governmental programs. 3 Although a key focus of prior research, elite capture is not the only reason why village leader selection for social programs may lead to sub-optimal targeting of beneficiaries. Notably, the dominant approach to evaluating the effectiveness of targeting is comparison of selected beneficiaries with the members of the community deemed the “poorest” according to some poverty indicator. This common metric for effectiveness in the research literature may differ from the metric in the mental model guiding the village leader selection process for several reasons. First, the set of the poorest will clearly depend on the indicator of poverty used, which is ultimately a subjective decision. Alatas et al. (2012) demonstrate that relatively worse performance of community leaders targeting compared to PMT stems from the fact that they use a different definition of poverty. Second, village leaders may use a set of selection criteria other than poverty indicators because they are optimizing for a different metric of effectiveness. For example, Basurto, Dupas and Robinson (2019) show that village chiefs target fertilizer subsidy distribution based on 2 Roopnaraine and Adato (2004), Brown et al. (2016) find that complexity of the PMT-based selection leads to challenges of acceptability since non-beneficiaries may find it difficult to understand why they were not selected while others were. Cameron and Shah (2014) find that using PMT in a cash transfer program in Indonesia led to distrust of local administrators, erosion of local social capital, and even an increase in crime rates. 3 By less than 1 percent. 2 expected productivity. The authors further demonstrate the presence of meaningful intra-village transfers, suggesting that in the presence of these transfers allocating fertilizer subsidy to more productive households may be a superior selection criterion compared to targeting based on poverty status. Even if village leaders were optimizing targeting of the poorest community members according to some objective definition of poverty, the effectiveness of their selection will be limited by the information available to them. Although local leaders are immersed in the life of the community, they may not have the same level of detailed information on every household that an enumerator sent to recover this information could retain. Although lack of information is a straightforward and universally plausible constraint on targeting efficiency there is a notable lack of research on the topic. To our knowledge, the only study that tests this information hypothesis is Alix-Garcia et al. (2019) in which the authors find that local leaders actually do provide accurate information on households, but acknowledge that the accuracy of their information may vary by context; for example, their information may be less accurate in larger and more dispersed communities. In this context, our paper attempts to make a threefold contribution to the literature on targeting by local leaders. First, we contribute to the body of evidence addressing traditional metrics of effectiveness of local leader targeting by answering two questions: i) do they select worse off community members? ii) How well does their selection compare to alternative approaches, including the canonical technical approach of PMT? Second, we attempt to shed light on the process of selection. Specifically, we explore what characteristics and criteria best rationalize local leader led beneficiary selection. This aspect of the paper contributes to the evidence that local leaders may use a different definition of poverty compared to program designer/policy maker. We test whether, in selecting the poorest, local leaders are more likely to rely on information about assets, likelihood of hunger, or subjective perception of immediate need. While assets are likely to capture long-term welfare, likelihood of hunger is a better proxy of current welfare, and perception of immediate need may reflect exposure to shocks. Third, we check whether local leaders rely on accurate information when making targeting decisions: i.e. whether what they know about household’s landholdings corresponds to landholdings as reported by the household. Another contribution of our paper is in analyzing the effectiveness of targeting by local leaders in the specific context of a public works program. We focus on a program in the Lao People’s Democratic Republic where village heads played an important role in selecting program beneficiaries. Constraints on effective targeting are likely to differ by program type. For example, the likelihood of nepotism may be higher for programs where there are no implicit participation costs (e.g. unconditional cash transfers, food vouchers), compared to programs with strings attached, such as public works programs, especially if working conditions are difficult. In contrast to the workfare setting of our study, much of the existing targeting literature is focused on conditional cash transfers, including work focused on the targeting of health insurance (Alatas et al., 2012), food subsidies (Alatas et al., 2012 and Basurto, Dupas and Robinson, 2019), and fertilizer subsidies (Basurto, Dupas and Robinson, 2019). We collect detailed welfare data on those who were identified as eligible to participate in the program by the village heads, as well as on a random sample of households from the same village. We also administer a series of ranking exercises to the village heads, aimed to elicit guiding principles for their selection, and to assess the quality of information they rely on. The paper is organized as follows: the next section describes the Road Maintenance Groups program, the process of beneficiary selection and provides 3 background on the position of village heads. Section 3 presents the study’s data. Section 4 analyzes whether the village heads indeed selected poorer households in the village. Section 5 explores what village heads prioritize when making selections. Section 6 analyzes what information they use. Section 7 explores the question of acceptability of the village heads’ selection for other community members. Section 8 concludes. 2. Road Maintenance Groups Program and Beneficiary Selection In rural Lao PDR, remote villages are connected to laterite highways via dirt access roads. Access roads are built by government-hired contractors but typically fall into extreme disrepair and often become unusable within a couple years due to a harsh rainy season. The Road Maintenance Group (RMG) program is a pilot program set up by Lao PDR Poverty Reduction Fund (PRF) 4 with the dual objectives of providing an income-earning opportunity to in-need households and extending the lives of the access roads. The RMG program trains participants to carry out road maintenance activities that do not require heavy machinery and subsequently employs them to provide basic road maintenance services throughout the year. PRF designed the program to be part-time. In order to support women’s empowerment, PRF only hires women to work in RMGs. The villages participating in the RMG program are poor and have few opportunities for wage work. Compared to rural households in the nationally representative sample in the Lao Expenditure and Consumption Survey (LECS) V, collected in 2013, RMG villages were poorer across a range of indicators including housing composition, ownership of durables, and nutrition. 5 Only 17% of working-age individuals in these villages had paid work outside the household at the start of the program. The RMG jobs were therefore highly attractive to populations living in these villages, but the program aimed to allocate the RMG spots to the poorest households in the villages. The village heads played an important role in this allocation. First, the village heads were responsible for publicizing the program and inviting households to the registration event, acting to some extent as gatekeepers for information about RMGs. Baird, MacIntosh and Ozler (2013) provide an example of informational elite capture. Their analysis of targeting effectiveness of Tanzania’s flagship CDD program suggests that wealth, education, access to media, and political engagement are positively correlated with awareness of the program, which is a necessary condition for benefitting from it. Such lack of access to information results in only mildly pro-poor targeting. Second, the village heads contributed to the selection of beneficiaries through updating their poverty ranking. Specifically, beneficiaries for the RMG program were selected through a two-step process in June 2018. In the first step, those eligible to participate in the program were identified. In the second step, as demand for the program exceeded the number of RMG jobs, a lottery (lucky draw) was carried out to 4 PRF is housed under the Government Office of Lao PDR. Its mission is to apply a community driven development approach to reduce poverty in Lao PDR, with a focus on rural populations, by improving infrastructure and access to services and resources. 5 Authors’ calculations using the baseline data for the RMG impact evaluation and LECS V. Available upon request. 4 divide all the eligible and interested into beneficiaries (RMG) and a waitlist (WL). 6 In this paper, we focus on the first step of the selection process. In order to qualify as ‘RMG-eligible’, registrants had to be at least 18 years old, female, and able-bodied. Women from poorer households were given priority based on the household poverty rank (poorest, poor, middle income and better off). Specifically, women in the bottom two poverty ranks (poorest and poor) were first granted entry. If the village quota for RMG-eligible was met, the process ended. If the quota was not met, women from the next poverty rank (middle-Income) were then entered. If the quota was still not met, women from the next poverty rank (better-off) were also entered. Eligibility was thus determined by poverty bracket; no distinction was made between women of the same bracket. The poverty status of each household was determined through a process that combined PMT, community- participation, and village-head updating. Household poverty ranks were originally constructed by PRF in 2016 based on PMT. In each village, there was a multi-day participatory process of PMT verification and adjustment by which households were assigned poverty ranks. During the registration process for the RMG program in 2018, village heads updated these rankings to account for changes in the two years since the original ranking. Twenty-four percent of households changed their poverty rank as a result of this updating. Figure 1 plots the original ranking against ranking by the village heads in 2018. Overall, 174 households changed poverty category. 7 Among them, 74 (10 percent) became less poor: moved from poorest to poor or middle income, from poor to middle income, or from middle income to better off. One hundred (14 percent) became poorer, moving from better off to middle income, poor and poorest, from middle income to poor and poorest, or from poor to poorest. If these movements reflect real changes in household welfare, reliance on village heads is efficiency improving. It would allow PRF to allocate the RMG program considering recent changes in household welfare, without carrying out another PMT test and community validation, which at the moment of the RMG program launch were prohibitively expensive. On the other hand, the change in poverty ranking of 24 percent of interested households may reflect nepotism, or as discussed above, be based on a definition of poverty different from the one preferred by PRF. A priori, we believe the former hypothesis to be more likely. First, village head is an elected position. Specifically, the village head is first elected by villagers, and then approved by district governor. The term lasts for 3 years, and village head can be re-elected. Notably, the villagers can report to the district management office, if they are not happy with the village head, and request re-election/re-appointment. Consequently, there are accountability mechanisms that create disincentives to engage in nepotism. Second, assessment of poverty status of the villagers is included in responsibilities of village heads. Decree on the Criteria for Poverty Graduation and Development outlines six criteria for graduating from poverty, and village heads should be receiving training on these criteria. These criteria are: (1) have safe and strong housing; (2) have assets and equipment necessary for their livelihoods and income generation; (3) have 6 PRF agreed to introduce randomized assignment of jobs with two objectives: (i) fair and transparent allocation of the jobs; (ii) possibility to carry out an impact evaluation of the RMG program. The data analyzed in this paper were collected as part of the IE. 7 We have the original 2016 poverty ranks for 716 households. For other 430 households, the rankings were not available. We are confirming the role of village heads in assigning these rankings. 5 labor, stable income or employment; (4) school age family members receive lower secondary school education; (5) have access to clean water and stable sources of energy; (6) have access to primary public health services. Decree suggests that implementing ministries should divulgate its provisions to local governments. The extent to which village heads adhere to these criteria when selecting beneficiaries in practice depends on several factors, such as success in disseminating the decree and village heads training on it, potential differences in interpretation of some of these criteria (e.g. stable employment) and village heads’ access to information about these criteria for all households in the village. Lastly, village heads may still select on a different criterion, as in Basurto, Dupas and Robinson (2019), especially given that there are no formal checks of accuracy in place. We study the outcomes of village heads selection process in 85 villages across 7 provinces in northern and southern Lao PDR where the RMG program was implemented. 8 3. Data To understand how village heads determined household eligibility and to assess the quality of their decisions in terms of targeting the poor, RMG baseline data collection included two targeting assessment components. First, in addition to collecting data on RMG and WL households that form the treatment and control groups for the impact evaluation, we administered the same survey to a Random Sample (RS) of village households with a woman of eligible age to participate in the RMG program (limited for simplicity in the random sample to ages 18-55). Households selected into the random sample could be replaced if they did not have a female household member of eligible age, if they had moved out of the village, if they could not be located after multiple attempts spanning at least two days, if they refused to participate, or if they were already part of the RMG or WL sample. 9 This sampling strategy enables comparison of household characteristics between the RMG-eligible and RS groups to establish whether the households granted entry to the lucky draw were truly among the villages’ poorer households. Second, we administered the village heads a targeting survey to probe the key factors influencing their selection of welfare beneficiaries and to understand their underlying perceptions regarding what constitutes poverty. 8 Although the RMG program was implemented in 86 villages, one village was dropped from the analysis because recruitment of beneficiaries there deviated from the model described above. 9 There was a total of 274 replacements: 151 households were replaced because they were already in the RMG/WL sample. The remaining replacements were due to: absence of women of eligible age in the household (36); households were absent due to harvesting season (20); households were no longer in the village (33); refusal (14) and other reasons, such as mental illness or temporary absence (20). 6 The data analyzed in this paper were collected as the baseline for the impact evaluation in September– October 2018. 10 Baseline data collection included household, individual, and targeting questionnaires. We administered the household questionnaire to the RMG, WL, and RS households. RMG and WL members completed the individual questionnaire in full, and, in RS households, a woman of eligible age to participate in the RMG program answered select sections. The household and individual questionnaires included the following modules: labor force participation, household income, housing conditions, durable possessions, land ownership, economic shocks, nutrition, and personal connections with the village heads. Our survey budget did not allow for inclusion of a consumption module. Under the baseline data collection, 1,888 households were surveyed: 333 RMG (treatment) households, 813 WL (control) households, and 742 RS households. For simplicity we will be referring to the groups of RMG or WL jointly as RMG-eligible. Each village head completed a targeting survey for a total of 85 respondents. The targeting questionnaire asked the village heads to rank 15 village households along a set of poverty-related indicators. The 15 households were comprised of 5 RMG, 5 WL, and 5 RS households, randomly chosen from the village’s pool of survey respondents. 11 To prevent fatigue from the multiple rankings, the women’s names were written on slips of paper, which the village head reordered for each ranking. The 85 village heads ranked 10 Results of program impacts on household income, nutrition, women’s empowerment, and other outcomes will be shared in a separate paper. 11 Sixteen villages did not have enough RMG, WL, and RS members to rank 15 individuals because of either smaller RMG-eligible memberships or participant absence at time of survey. In these cases, the maximum number of households available for ranking was specified in the instructions. Eight villages had 14 households ranked, 5 villages had 13, 2 villages had 12, and 1 village had 9. 7 a total of 1,245 households. There were two types of ranking questions: RMG priority ranking and ranking based on specific criteria: “RMG Priority Ranking”: Please rank who you would prioritize to join the RMG if it were solely up to you and not determined by eligibility and a lottery. Start with the highest priority to the least priority. For example, if I were in the list and you really thought I should receive the RMG job, you would rank me number 1. If you think I am the last person in the list you would give the job to, rank me (total number in the list). 12 Criteria Rankings: • Rank who owns the least to most square meters of land, agricultural and residential land combined. • Rank the likelihood that, in the last 12 months, the person or someone in her household was hungry but did not eat because there was not enough rice or food at home, and not enough money or other tradeable goods to purchase rice or food. Start with the person or household most likely to have had been hungry but did not eat. • Rank who needs the most to least help right now. The variety of data collected, including actual data on selection, respondent self-reports, and data collected on hypothetical ranking by village heads, combined with an intention sampling design that includes both selected and non-selected among eligible and a random sample of villagers, enable us to explore both the actual targeting by village heads, the proximate inputs into the targeting choices, and the potential information constraints that may limit the ability of village heads to target the poorest households in their villages. 4. Do village heads effectively select poorer households for participation in RMGs? Given the extensive baseline surveys administered to the RMG-eligible and the RS, we can compare them on a wide range of dimensions, including income, assets, land holding, etc. Notably, the baseline questionnaires did not include a comprehensive consumption module, which elsewhere is commonly used for assessing poverty status. Table A1 in the Appendix shows 127 variables describing the participants’ basic demographic characteristics, household composition, income, possession of durables and livestock, housing conditions, nutrition, and experience of economic shocks, available in our survey. To reduce dimensionality and address the likelihood of Type I errors in our numerous possible comparisons between the RMG-eligible and RS groups, we follow Katz, Kling and Liebman (2007) and Andersen (2008) by constructing composite z-scores and adjusting the p-values for multiple hypothesis testing. The z-scores of the individual variables are calculated and then averaged by category to construct composite z-scores. The z-scores of variables assumed to be negatively correlated with wealth are inverted. Altogether, we construct 7 composite z-scores to encompass broad dimensions of welfare: land and buildings ownership, household composition, durables ownership, livestock ownership, housing conditions, nutrition, and shocks. We also test for differences in household per capita total income. We apply Bonferroni correction to adjust p-values for the differences in means tests, to further reduce the 12 Sixteen villages did not have enough RMG, WL, and RS members to rank 15 individuals because of either smaller RMG-eligible memberships or participant absence at time of survey. In these cases, the maximum number of households available for ranking was specified in the instructions. Eight villages had 14 households ranked, 5 villages had 13, 2 villages had 12, and 1 village had 9. 8 likelihood of false rejections of the null hypothesis of no difference due to multiple hypotheses testing. Table 1 shows the t-tests and provides more details on variables included in specific z-scores. Table 1: Differences in household welfare between RMG-eligible and Random Sample Diff. (RMG- Bonferroni RMG- Random RMG- Random eligible – corrected eligible Sample eligible Sample P-val Random obs. obs. mean mean P-val Sample) Land and buildings 1146 742 -0.035 0.054 -0.0890** 0.004 0.033 ownership Household composition 1146 742 -0.0176 0.0272 -0.0448 0.013 0.102 Durables ownership 1146 742 -0.0388 0.0599 -0.0987*** 0.000 0.000 Livestock ownership 1146 742 -0.0449 0.0693 -0.1142*** 0.000 0.001 Total income per HH 1146 742 381.919 480.733 -98.8140** 0.004 0.032 member (USD PPP) Housing conditions 1146 742 -0.0131 0.0202 -0.0333* 0.008 0.066 Nutrition 1146 742 0.0097 -0.0149 0.0246 0.084 0.669 Exposure to shocks 1146 742 0.0039 -0.0059 0.0098 0.666 1.000 * p<0.1 ** p<0.05 *** p<0.01 RMG-eligible = 1146 observations. Random Sample = 742 observations. All household characteristics are shared as composite z-scores, except for total income per HH member. Variable included in composite z-scores are: Land and buildings: total area of residential and agricultural land, plus total number of buildings owned by each household. Household composition: dependency ratio, household head age, gender, ethnicity, and education. Durables ownership: ownership of household goods, including cars, TVs, refrigerators, etc. Livestock ownership: numbers of productive animals owned, including cows, buffalos, chickens, etc. Total income per capita: sum of 9 types of income divided by the number of household members. Housing conditions: number of rooms; flooring, roofing, and wall materials; kitchen location and structure; access to water; etc. Nutrition: balls of rice, kg of meat, number of days dairy eaten, etc. Exposure to shocks: experience of health, business, and agricultural shocks. Table 1 suggests that the RS households are significantly better off than RMG-eligible households, typically owning more possessions, having higher quality housing, and generally experiencing better well-being as viewed through these variables. It is important to remember that the pool of RMG-eligible reflects a combination of targeting methods: self-targeting as well as targeting by village heads. Village heads affected the selection process through (i) sharing information about the program; (ii) updating poverty 9 status. 13 However, only women who were also interested to join RMGs entered the lucky draw and are captured in our sample as RMG-eligible. For simplicity and clarity of exposition, we may refer to this sample as village head selection, though formally speaking this implies village head selection conditional on self-selection. Overall, Table 1 suggests that the RMG program effectively targeted poorer households within villages. We next compare actual selection to a hypothetical selection based on PMT used by the Government of Lao PDR to identify the poor for participation in social programs. This PMT index was developed by the World Bank and Government of Lao PDR, following administration of LECS V in 2012-2013 (Pimhidzai et al. 2014). It relied on regression analysis of household consumption on surveyed household characteristics. The model aimed to accurately distinguish between poor and non-poor households. Non- significant characteristics were removed from the model; characteristics that minimized misidentification of poor and non-poor households were kept. Index variables and weights were determined based on the regression outcomes. Households scoring below 64 were classified as poor. Table A2 in the Appendix provides the complete list of variables used in the PMT index (World Bank forthcoming). When we compare village head selection with those selected through PMT, we find that, though both methodologies selected approximately the same number of beneficiary households, they disagree on which households should benefit in 44% of cases. The magnitude of this disagreement holds true whether one considers the full sample of 1,888 study households (Table 2) or only 1,245 households ranked in the village head targeting survey (Table A3 in the Appendix). Although we expect differences in how PMT and village heads identify poor households, the consensus on population poverty rates and need for welfare support is striking. Table 2: Village heads vs PMT – Full Sample of 1,888 Households PMT below 64 RMG-eligible Total No Yes No 300 (16%) 442 (23%) 742 (39%) Yes 401 (21%) 745 (39%) 1,146 (61%) Total 701 (37%) 1,187 (63%) 1,888 (100%) The discrepancy in identification of the poor by the two methods leads us to further assess their targeting accuracy. We compare household characteristics typically indicative of wealth status between the eligible and ineligible/ poor and non-poor groups under village heads selection and PMT. We expect the more accurate method to produce a larger difference in observable welfare characteristics. To reduce the complexity of analyzing overall wealth disparities through numerous household characteristics, we again construct composite z-scores similar to those we used for comparisons of RMG- 13 Thus, the selection may also reflect PMT and community validation carried out in 2016. However, based on the field supervisors’ reports, the changes in poverty ranks were substantial. 10 eligible and random sample in Table 1. One important difference in the construction of these z-scores is that now we exclude variables used to construct the PMT index. Including these variables would give unfair advantage to the PMT method: we would expect to observe larger differences between eligible and non-eligible according to PMT on the variables underlying its construction. 14 Figure 1 plots the differences in z-scores between RMG-eligible and RS, against the differences between poor and non-poor according to PMT, including 90% confidence intervals. The gaps between groups according to the two methods suggest that PMT outperforms RMG-eligible selection in identifying poor households. The differences in means are significantly larger under PMT for 4 out of 8 analyzed indicators. 15 RMG-eligible selection performs better than pure random assignment to eligible vs. ineligible groups, but it is less accurate than PMT. T-tests for differences in z-scores for eligible and not eligible according to two selection methods are shown in Table A3 of the Appendix. However, village heads were identifying the poorest households from within a village, not across the sample. If we replicate this analysis by calculating z-scores within village, rather than across the entire sample, village head selection and PMT appear to provide equivalent levels of accuracy. Figure 2 shows the differences in z-scores between eligible and non-eligible according to the two methods, including confidence intervals. 16 The overlap of confidence intervals suggests that we cannot reject the hypothesis of no difference in the magnitude of the welfare gap between eligible and non-eligible according to the two methods. Table A4 in the Appendix shows corresponding t-tests. In the analysis based on village-level z-scores, easily observed variables perform better under village heads selection, while factors less visible to the public eye perform better under PMT. For example, livestock and durables ownership are negative and significant for both methodologies, but the difference in means is greater under RMG-eligible selection. Conversely, nutrition and income, the status of which might be known only to household members, perform better under PMT; the difference in the nutrition composite z-score t-test is negative and significant only for PMT, and the difference in means for income is significantly larger under PMT. The village-level results suggest that when organizing beneficiary selection to meet the needs of the local context, the village heads selection was effective through leveraging the heads’ familiarity with the households. This familiarity, however, was limited when it came to assessing more obscure or private household characteristics. In such cases, the rigorous PMT methodology based on detailed survey data performed better. We must also note that less-poor households (ranked ‘Middle-income’ or ‘Better-off’) were included in the RMG-eligible list in some villages in order to satisfy village quotas for the program. We would expect the inclusion of these non-poor households to worsen the average performance of the village heads selection. In a setting where the quotas for road maintenance were satisfied by the households from the 14 Comparison of RMG eligible observations and random sample using this restricted selection of variables provides results similar to those shown in Table 1. See Table A5 in the Appendix for all comparisons. 15 Please note, income per household member is shown as z-scores in Figure 1 in order to display the measure along with the other indicators in the graph, but the actual difference for the RMG-eligible vs. RS is -0.1426 (p<0.01) and, for PMT poor vs. nonpoor, -0.8213 (p<0.01). 16 Income per household member is again shown as a z-score to aid the graph. Actual difference between RMG- eligible and RS is -0.2160 (p<0.01). Between PMT poor and nonpoor, the difference is -0.5417 (p<0.01). 11 lowest poverty ranks, we would expect the village heads selection to fare slightly better on welfare comparisons. 12 Figure 1: T-tests of composite z-scores by VHT and PMT targeting methods – Across the full sample Figure 2: T-tests of composite z-scores by VHT and PMT targeting methods – Grouped by village 5. How do village heads prioritize potential beneficiaries for RMG selection? We next attempt to understand the mechanism behind village heads’ selections. To do so, we use the ranking exercises we administered to the village heads. The objective of asking the village heads to choose among the pooled list of RMG-eligible and randomly selected households was two-fold: First, we check whether village heads’ views on who should be working in RMGs are aligned with PRF- imposed targeting criteria. It is conceivable that village chiefs may disagree with making household poverty level a core criterion for selecting RMG participants. For instance, village heads may be concerned with the quality of public goods delivered and prioritize women’s ability to do the job. To achieve this, we 13 emphasize “if it were solely up to you and not determined by eligibility and a lottery” when asking village heads to rank households in order of priority for RMG jobs. Second, we aim to understand the driving factors behind the village heads’ selection. The ranking exercise helps us in several ways. First, we can compare RMG priority ranking with rankings on three criteria, which broadly capture different potential approaches to prioritizing the poor: ranking on agricultural and residential land, on likelihood to experience hunger, and on the need of help. Ranking on assets is likely to capture long-term poverty status, ranking on the likelihood of hunger may potentially capture greater severity of poverty, and ranking on the need of help is likely to capture exposure to shocks. Correlating RMG priority ranking with rankings on different dimensions of poverty will provide information on which dimensions of need village heads prioritize when selecting for RMGs. As Figure 3 makes clear, the RMG priority ranking is largely aligned with the village heads’ selection of households eligible to enter the RMG job lucky draw. In Figure 3, the x-axis shows the RMG priority ranking and the y-axis shows the percentage of RMG-eligible and RS households occupying each spot in the priority ranking. For example, among all observations ranked as #1 priority, slightly more than 90% are from the RMG-eligible group and less than 10% are from the RS. At least 80% of the first 10 priority ranking spots are filled by RMG-eligible households, typically leaving the last 5 spots in the priority ranking for the 5 RS households included in the ranking exercise. Figure 3: Fraction of RMG-eligible and Random Sample in each rank position Although the fact that village heads assign RMG-eligible households to the highest positions in the RMG- priority ranking may suggest their agreement with the principle of selection into RMGs based on poverty status, it is important to note that there may also be an anchoring effect. We conducted the ranking exercises shortly after actual selection into RMGs took place. The village heads may have been inclined to select women they already knew to be in the RMG-eligible group. We next proceed to exploring how the three criteria rankings (land holdings, likelihood of being hungry, and need for help) correlate with the RMG priority ranking. Does a higher ranking among one of these criteria create a greater likelihood of being ranked higher in the RMG priority ranking? Table 3 14 demonstrates a regression of likelihood of being ranked higher for RMG priority on each of the three criteria in columns (1) – (3) and the three rankings together in column (4). Regression in column (5) controls for RMG-eligibility status: a dummy equal to 1 if the household is RMG-eligible, and 0 if it belongs to the random sample. Table 3 reveals several interesting features about the process of village heads’ selection. First, the relationship between priority ranking and each of the three ranking is significant, but coefficients are low, ranging between 0.15 (for land holdings) and 0.22 (for the need for help). Notably, association between the ranking of those who need help most and RMG priority ranking is the strongest: not only is the coefficient the highest, association between these rankings is the only one that remains significant once all three rankings are accounted for in a regression. Village heads appear to rely on some notion of the need of help when deciding on how to prioritize for RMGs. Second, once we account for RMG-eligibility, none of the three criteria remains significant. Only designation as an RMG-eligible or RS household is significant: being eligible for RMGs is associated with decrease in rank, and consequently higher priority for RMGs in village heads’ perception (please note that households are ordered from 1 to 15, one being the highest priority). This may suggest that village heads fully endorse the selection criterion proposed by PRF: household’s poverty status. An opportunity to prioritize based on a different criterion does not lead to a vastly different result than their selection of RMG-eligible. However, as mentioned earlier, we cannot exclude the possibility of an anchoring effect. Table 3: Likelihood of being prioritized for an RMG role as a function of other rankings Priority for Priority for Priority for Priority for Priority for RMG RMG RMG RMG RMG Lowest to highest land holdings 0.1560*** 0.0465 0.0266 (0.036) (0.038) (0.028) Highest to lowest likelihood to be 0.1834*** 0.0321 -0.0277 hungry (0.036) (0.047) (0.037) Highest to lowest need for help 0.2228*** 0.1751*** 0.0466 (0.037) (0.049) (0.040) RMG-eligible (y/n) -6.2117*** (0.420) Village fixed effects Yes Yes Yes Yes Yes Observations 1,245 1,245 1,245 1,245 1,245 * p<0.1 ** p<0.05 *** p<0.01 Priority for RMG regressions are calculated as highest to lowest priority: the household ranked #1 is the household that should get the highest priority. Standard errors in parentheses. The 3 criteria (land holdings, likelihood to be hungry, need for help) are ranked with #1 as the poorest or worst outcome, and improving in status as the rankings increase until the final slot, #15. 6. What information do village heads use in their targeting decisions? Relatively weak association between RMG priority ranking and the three rankings which may capture different dimensions of poverty begs the question of whether village heads use accurate information in assigning the ranks. To check whether this is the case, we correlate village heads’ ranking and ranking 15 based on the survey data. Specifically, we correlate ranking on land with rank on the sum of land area, ranking on the likelihood of hunger with a nutrition index, and ranking on the need for help with an index constructed based on the survey data about recently experienced shocks. We use Kendall correlations, a special case of standard Pearson correlation evaluating the degree of similarity among sets of ranks (Abdi, 2007). To check robustness, we use several measures for each of the dimensions of poverty captured in ranking exercise. For land, in addition to the sum of residential and agricultural plots, we create a composite z- score of variables related to land and building ownership, as well as the first principal component of these variables. For the likelihood of being hungry, we use a composite z-score and the first principal component of all nutrition-related questions in the survey. For the likelihood of needing help, we create a composite z-score of all survey variables related to economic shocks, and also use the first principal component of these variables. These variables provide a unique household ranking in nearly all villages. Table 4 presents the results. Village heads’ rankings of land ownership are weakly but significantly correlated with households’ reported land size. Their rankings of likelihood of hunger are correlated with nutrition measures but the correlation is weaker. Rankings of need for help are not correlated with shocks. The consistent strength of land measures across ranking exercises may be explained by the fact that land is the most directly observable of the three criteria for someone outside the household. Table 4: Correlations of rankings and household variables Lowest to highest Highest to lowest Highest to lowest land holdings likelihood to be need for help hungry Land measures Size of all plots (res. and agr.) 0.149*** 0.098*** 0.103*** Composite z-score of land and buildings 0.150*** 0.133*** 0.136*** Principal component 1 of land and buildings 0.148*** 0.133*** 0.137*** Nutrition measures Composite z-score of nutrition 0.008 0.004 0.017 Principal component 1 of nutrition 0.088*** 0.108*** 0.090*** Shocks measures Composite z-score of shocks (all inverted) -0.049** -0.031 -0.023 Principal component 1 of shocks (all inverted) -0.031 -0.028 -0.008 Observations 1,245 * p<0.1 ** p<0.05 *** p<0.01 The weakness of the correlations between the three criteria and their corresponding observable characteristics indicates that village heads do not determine these rankings based solely on observable land, nutrition, and shock characteristics. Rather, the village heads may base their rankings on other information about household welfare that overlaps with the information on land, nutrition, and shocks. 16 The question is: what are the indicators that village heads use to decide which households have more land, are hungrier or more in need of help? Random forest analysis facilitates this inquiry by ranking independent variables—welfare characteristics in our case—from strongest to weakest, in terms of their explanatory power to predict the dependent variable outcome—here, likelihood of being ranked highly in one of the ranking exercises (Friedman et al., 2001). It is important to note that random forest analysis determines the strength of the association, but not its direction. In random forest regressions, the predictive power of a variable shows the extent to which a given variable, when used to predict the values of the dependent variable, contributes to the increase in the accuracy of predictions. In all four random forest graphs presented below, average increases in accuracy of predictions and the importance of a given variable are expressed relative to the most important variable, which has a value of 1. In Figure 4, for example, we can see that “Woman works in non-HH business” is the most important variable for predicting how high she is prioritized for an RMG role. An external unroofed kitchen is the next most important, with a value of approximately 0.725, roughly ¾ as predictive as whether the woman works outside the household. The scale, however, does not indicate if working outside the household makes the woman more likely to be prioritized—according to the scale, it could heavily weigh against her, as well. Figure 4: Random forest analysis of RMG priority ranking Figure 4 demonstrates that the most predictive indicators for RMG priority ranking are easily observable household characteristics, such as the woman working in a non-household business, possessing an outside unroofed kitchen, and having concrete dwelling walls. Factoring in women’s work in non-household businesses might signify that the village head considered women’s need for work or capability to maintain a role outside the home. Income from services is also a predictor, potentially indicating that village heads took monetary well-being into account when prioritizing RMG members. Figure 4 displays only the 20 17 strongest predictive indicators for RMG priority ranking; the complete list of 127 predictors is shared in the Appendix. Figures 5-7 show the 20 most predictive indicators for each of the three criteria. Similar to the RMG priority ranking, the main predictors for the three criteria are easily observable household characteristics, plus total income. Although total income is predictive, its highest ranking is 4, for likelihood of hunger, implying that directly observable characteristics matter more than income for village heads’ ranking of beneficiaries. Figure 5: Random forest analysis of land holdings ranking The three targeting survey criteria rankings largely share predictive indicators, though the ranking order varies between the criteria. The following 10 indicators are among the top five predictors for one or more of the three ranking criteria, and, aside from bamboo walls and female household heads, their presence likely signifies wealthier households: (i) tile dwelling floors; (ii) concrete dwelling walls; (iii) bamboo dwelling walls; (iv) female household head; (v) number of dwelling rooms; (vi) number of mobile phones; (vii) number of motorcycles; (viii) income from services; (ix) total income; (x) woman works in household business. Most of these indicators are directly observable. Interestingly, the independent variables on land holdings do not fall among the top 10 predictors for any of the criteria, including the village heads’ ranking of land ownership. This suggests that, while village heads select on observable characteristics, these are not necessarily the characteristics most directly linked even to criteria they are ostensibly ranking. It is important to note that baseline data collection ran a few months after the RMGs had been formed and RMG work had begun, so it is possible that the RMG-eligible group reported more conversations with the village heads because they were speaking with the village heads about RMG work. However, it is also possible that there was some elite capture. We must also note that these two variables were self-reported by the survey participants, and possibility for error through misunderstanding or misrepresentation exists. 18 Taken at face value, though, the variables indicate minimal or no elite capture, nepotism, or favoritism in the village head targeting. Figure 6: Random forest analysis of the likelihood of being hungry ranking Figure 7: Random forest analysis of the need for help ranking Random forest analysis suggests that the village heads’ ranking of households on specific criteria is based on a holistic impression of poverty gathered from a host of observable characteristics. This holistic assessment has precedence. Alatas et al. (2012) write that community representatives assigned to select program beneficiaries may employ holistic measures of well-being that account for local context, including 19 context-specific understandings of poverty and of who needs help the most. The holistic measures are based on visible household characteristics and potentially on observable community characteristics, as well. 7. Is the village head’s role in selection acceptable to the population? While discussing the random forest analysis of ranking predictors, we observed that there appears to be little or no evidence of elite capture in the village head selection process. Analyzing the performance of village head targeting against PMT suggested that the village heads are better able to incorporate local knowledge, tailoring beneficiary selection to the local context. To access effectiveness of village heads targeting, one more question must be answered: do community members consider village head targeting to be a legitimate method for allocating community resources? Our survey included two questions that facilitate this inquiry: 1) Do you feel that your village head has a good understanding of how rich or poor most of the households are in your village? 2) When the RMG in this village was selected, the village head advised the selection team on the poverty status of the applicants. Is this a fair practice? These questions were administered to WL and RMG households only – unfortunately, we did not ask the RS households. Below we present the results separately for RMG and WL women, considering that being selected into/out of the program, even if as a result of lottery, may affect RMG/WL women’s view on the overall fairness of the selection process and the role of village heads. Both RMG and WL women indicate strong confidence in the accuracy and fairness of village head targeting. Sixty-five % of RMG and 67% of WL women thought the village head “absolutely” had a good understanding of the villagers’ wealth. An additional 22% of WL and 19% of RMG members thought the village head understood “to some degree.” There was no significant difference between their responses. Similarly, the RMG and WL women considered the village head’s role in beneficiary selection to be fair. Sixty-eight % of WL and 70% of RMG thought it was “absolutely” fair, and 18% of WL and 17% of RMG thought it was fair “to some degree.” Again, there was no significant difference in answers between WL and RMG. Unfortunately, we did not ask the RS households these questions. This leaves open the possibility that those who were admitted to the lucky draw were satisfied with village heads involvement in the process even if they did not ultimately benefit from the program (i.e. were in the control group), whereas those denied the chance to enter the lucky draw could be dissatisfied with the village heads’ role. It is also possible that our respondents indicated satisfaction out of acquiescence bias or hesitation to admonish authority figures. Future studies can inquire more deeply into community-wide acceptance of village head targeting. 8. Conclusion We studied the targeting performance of a multi-province workfare program in rural Lao PDR that used a combination of methods to target poorer women interested in the opportunity to do paid work. To identify poorer women, the program used poverty ranks constructed using PMT and community ranking in 2016 and updated by village heads during the program registration process in 2018. Our results suggest that the program was successful in identifying beneficiaries significantly worse off than a random sample 20 of households from their villages. Moreover, comparison with a hypothetical pure-PMT demonstrates the two methods to be roughly equivalent in their ability to identify poorer households within a village. Further, we shed light on what information village heads use to assess the welfare status of households. We asked village heads to rank households by different welfare criteria and then compared these data to household survey data. We find that village heads seem to base their assessments of specific criteria like household land ownership, likelihood of being hungry and likelihood of needing assistance on a holistic impression of household welfare, rather than specific indicators like actual land ownership, nutrition or economic shocks. Using random forest analysis, we find that their rankings are best explained by easily observable indicators of welfare like household characteristics and asset ownership. In this program, updating poverty ranks by village heads provided a much quicker and less expensive alternative to a full PMT plus community-ranking and validation process, without compromising the targeting efficacy of the program and produced outcomes acceptable for the community. The results suggest that the performance of village heads compared to alternative methods may be particularly strong in remote areas with varying poverty levels. There, the village heads’ familiarity with residents and local circumstances can help them distinguish among community members’ relative welfare statuses. In these cases, a flexible targeting led by village heads may be successful at achieving pro-poor targeting objectives. There are caveats about the applicability of these findings in other contexts. First, village heads only updated prior, community-supported poverty rankings in this program. This reduced the reliance on village heads for determination of eligibility and may have also improved community acceptability, since applicants may have perceived the updated rankings as a revision of a system they had already endorsed. Other populations without a comparable experience may view village head targeting as less legitimate. Second, the selection for the RMG program also relied on self-selection into the program by the neediest households. This may explain some of the targeting success of the program. Third, our analysis suggests that village heads hold a holistic impression of household welfare built on easily observable household and individual characteristics. This type of assessment may be less useful when targeting households with specific, less observable needs, such as the need for nutrition support. 21 References 1. Abdi, H. (2007). The Kendall rank correlation coefficient. Encyclopedia of Measurement and Statistics. Sage, Thousand Oaks, CA, pages 508–510. 2. Alatas, V., Banerjee, A., Hanna, R., Olken, B. A., and Tobias, J. (2012). Targeting the poor: evidence from an experiment in Indonesia. American Economic Review, 102(4):1206–40. 3. Alix-Garcia, J.M., Sims, K.R.E., and Costica, L. (2019). Better to be indirect? Testing the accuracy and cost-savings of community leader surveys for poverty targeting. Working paper, University of California Davis. 4. Baird, Sarah, Craig McIntosh and Berk Ozler (2013) The Regressive Demands of Demand-Driven Development, Journal of Public Economics, Vol. 106, pp. 27-41. 5. Basurto, M. P., Dupas, P., & Robinson, J. (2019). Decentralization and efficiency of subsidy targeting: Evidence from chiefs in rural Malawi. Journal of Public Economics, 4047. 6. Brown, C., Ravallion, M., and Van de Walle, D. (2016). A poor means test? Econometric targeting in Africa. Technical report, The World Bank, Washington, DC. 7. Cameron, L. and Shah, M. (2014). Mistargeting of cash transfers, social capital destruction, and crime in Indonesia. Economic Development and Cultural Change, 62(2):381– 415. 8. Conning, J. and Kevane, M. (2002). Community-based targeting mechanisms for social safety nets: A critical review. World development, 30(3):375–394. 9. Friedman, J., Hastie, T., and Tibshirani, R. (2001). The elements of statistical learning, Volume 1. Springer series in statistics: New York. 10. Kilic, T., Whitney, E., Winters, P., 2013. Decentralized beneficiary targeting in largescale development programs: insights from the Malawi farm input subsidy program. Journal of African Economies, Volume 24, Issue 1, January 2015, Pages 26–56. 11. Kling, J. R., Liebman, J. B., & Katz, L. F. (2007). Experimental analysis of neighborhood effects. Econometrica, 75(1), 83-119. 12. Pimhidzai, O., Fenton, N. C., Souksavath, P., and Sisoulath, V. (2014). Poverty Profile in Lao PDR. Technical report, The World Bank. 13. Roopnaraine, T. and Adato, M. (2004). A social analysis of the Red de proteccion social (RPS) in Nicaragua. Washington, DC: IFPRI. 14. Subbarao, K., Del Ninno, C., Andrews, C., and Rodríguez, C. (2012). Public works as a safety net: design, evidence, and implementation. The World Bank. 22 Appendix Table A1: T-tests of RMG-eligible and RS Group Descriptive Statistics Note: Yellow highlighting indicates that the variable was excluded during construction of composite z-scores to compare VHT against PMT, since these variables were used in construction of the PMT index. Note: Variables marked "(inverted)" were not inverted in Table A1, but their individual z-scores were inverted when computing composite z-scores. Diff. (RMG- RMG-eligible Random Random RMG-eligible eligible – mean Sample mean Sample P-val SD Random (N=1,146) (N=742) SD Sample) Land and buildings ownership Size of agr. plots (thousand sqm) 19.3528 21.4625 19.6314 24.212 -0.2785 0.799 Size of res. plots (thousand sqm) 0.3287 0.6168 0.4269 0.7434 -0.0982*** 0.003 Num. of buildings owned (res. and business) 1.0785 0.2969 1.1146 0.3805 -0.0360** 0.029 Household composition HH dependency ratio (inverted) 0.9138 0.6953 0.8761 0.7146 0.0377 0.258 Head of HH age (years) 41.5175 12.8011 43.6739 13.2874 -2.1564*** 0.000 Head of HH education (years) 3.2818 2.9826 3.5957 3.3598 -0.3138** 0.039 Head of HH is female (inverted) 0.0925 0.2899 0.0472 0.2121 0.0453*** 0.000 Head of HH is divorced (inverted) 0.0244 0.1545 0.0175 0.1313 0.0069 0.298 Head of HH is a widow (inverted) 0.0742 0.2622 0.0418 0.2002 0.0324*** 0.002 Head of HH is married 0.8988 0.3018 0.9407 0.2363 -0.0419*** 0.001 Head of HH was never married (inverted) 0.0026 0.0511 0 0 0.0026* 0.083 Head of HH ethnicity: Lao 0.1309 0.3374 0.1173 0.3219 0.0136 0.378 Head of HH ethnicity: Khmu 0.4319 0.4956 0.3733 0.484 0.0586** 0.011 Head of HH ethnicity: Katang (inverted) 0.1169 0.3215 0.1658 0.3721 -0.0488*** 0.003 Head of HH ethnicity: Other (inverted) 0.3202 0.4668 0.3437 0.4753 -0.0234 0.292 Durables ownership Num. of vehicles (car, vans, etc.) 0.0253 0.1626 0.0499 0.2413 -0.0246** 0.015 Num. of motor cycles 0.7452 0.7451 0.8814 0.7789 -0.1362*** 0.000 Num. of bicycles 0.0497 0.233 0.066 0.2933 -0.0163 0.202 Num. of refrigerators/freezers 0.0908 0.3272 0.1685 0.4346 -0.0777*** 0.000 Num. of sewing machines 0.055 0.228 0.0714 0.2577 -0.0165 0.157 23 Num. of washing machines 0.0035 0.059 0.0216 0.1454 -0.0181*** 0.001 Num. of electric rice cookers 0.0628 0.2567 0.097 0.3265 -0.0342** 0.016 Num. of steam rice cookers 1.4572 0.645 1.3989 0.6341 0.0583* 0.053 Num. of food processors 0.007 0.0833 0.0202 0.1501 -0.0132** 0.029 Num. of two- and four-wheel tractors 0.2373 0.4398 0.3369 0.4898 -0.0996*** 0.000 Num. of agr. equipment 9.6152 5.909 10 6.3484 -0.3848 0.187 Num. of boats 0.0175 0.131 0.0377 0.1976 -0.0203** 0.014 Num. of fishing nets 0.548 0.8836 0.6402 1.0147 -0.0922** 0.043 Num. of rice mills 0.2784 0.4523 0.3113 0.4662 -0.033 0.129 Num. of TVs 0.3229 0.4807 0.4205 0.5334 -0.0976*** 0.000 Num. of radios/VCD 0.1457 0.3928 0.1752 0.4078 -0.0295 0.120 Num. of telephones 0.1178 0.3609 0.1226 0.3482 -0.0048 0.771 Num. of mobile phones 1.4145 1.395 1.4569 1.3667 -0.0424 0.514 Num. of satellite discs/connections 0.3054 0.4608 0.3814 0.4861 -0.0760*** 0.001 Num. of computers 0.0087 0.093 0.0216 0.1544 -0.0128** 0.042 Num. of air conditioners 0.0009 0.0295 0.0027 0.0519 -0.0018 0.384 Num. of jewelry 0.301 0.854 0.3383 0.9524 -0.0372 0.388 Num. of mosquito nets 2.9206 1.7012 3.0418 1.8323 -0.1212 0.149 Num. of solar panels 0.2286 0.4617 0.2776 0.5653 -0.0490** 0.049 Livestock ownership Num. of cows 1.1745 2.7296 1.973 3.896 -0.7985*** 0.000 Num. of buffalos 0.5567 1.3927 0.9151 2.0999 -0.3584*** 0.000 Num. of goats 0.5192 1.838 0.8518 2.2621 -0.3326*** 0.001 Num. of pigss 1.9188 2.464 2.0526 2.7296 -0.1337 0.280 Num. of chickens 9.7164 14.5332 9.9447 14.0601 -0.2283 0.734 Num. of ducks 2.1195 4.4355 2.1132 4.4888 0.0063 0.976 Income Agriculture income 382.5096 1035.4527 406.5878 1085.7145 -24.0782 0.632 Livestock income 390.1721 916.5445 513.3494 1135.7117 -123.1773** 0.013 Fishes income 6.4066 26.6132 7.2247 29.5335 -0.8181 0.541 Forest income 159.6157 307.38 137.9027 323.628 21.713 0.147 24 Handicraft income 94.7048 340.7649 89.1626 360.8131 5.5421 0.739 Services income 30.8876 181.66 49.3024 232.1116 -18.4148* 0.068 - Salary income 435.9056 1741.3517 809.0677 2522.6435 0.000 373.1622*** Extra income 150.4058 520.9682 114.4544 448.0776 35.9515 0.111 Other income 114.9503 380.2719 102.258 360.7227 12.6924 0.465 - Total income 1933.775 2951.829 2410.1417 3614.1872 0.003 476.3667*** Total income per HH member (USD PPP) 381.919 620.3777 480.733 789.1612 -98.8140*** 0.004 Housing conditions Dwelling: num. of rooms 1.9171 1.1149 2.0782 1.1176 -0.1611*** 0.002 Dwelling walls: brick 0.0654 0.2474 0.0755 0.2643 -0.01 0.409 Dwelling walls: concrete 0.0637 0.2443 0.0903 0.2868 -0.0266** 0.037 Dwelling walls: wood 0.5663 0.4958 0.624 0.4847 -0.0577** 0.012 Dwelling walls: bamboo (inverted) 0.2862 0.4522 0.1873 0.3904 0.0989*** 0.000 Dwelling roof: concrete 0.0052 0.0722 0.0013 0.0367 0.0039 0.123 Dwelling roof: wood 0.0271 0.1623 0.0216 0.1454 0.0055 0.444 Dwelling roof: metal (inverted) 0.5977 0.4906 0.6173 0.4864 -0.0195 0.396 Dwelling roof: tile 0.3028 0.4597 0.3167 0.4655 -0.0139 0.524 Dwelling roof: grass (inverted) 0.0611 0.2396 0.0418 0.2002 0.0193* 0.059 Dwelling floor: tile 0.1213 0.3266 0.1402 0.3474 -0.0189 0.238 Dwelling floor: concrete 0.0733 0.2607 0.0768 0.2665 -0.0035 0.777 Dwelling floor: wood (inverted) 0.5672 0.4957 0.6011 0.49 -0.0339 0.144 Dwelling floor: bamboo (inverted) 0.0864 0.2811 0.0606 0.2388 0.0257** 0.033 Dwelling floor: earth/clay (inverted) 0.1518 0.359 0.1213 0.3267 0.0305* 0.057 Rainy season water: piped water 0.178 0.3827 0.1752 0.3804 0.0028 0.876 Rainy season water: protected well 0.5332 0.4991 0.4704 0.4995 0.0628*** 0.008 Rainy season water: unprotected well (inverted) 0.0515 0.2211 0.0687 0.2532 -0.0172 0.129 Rainy season water: river/lake (inverted) 0.1344 0.3412 0.1132 0.3171 0.0212 0.169 Dry season water: piped water 0.1806 0.3849 0.1792 0.3838 0.0014 0.939 Dry season water: protected well 0.5663 0.4958 0.5162 0.5001 0.0501** 0.033 25 Dry season water: unprotected well (inverted) 0.0812 0.2732 0.1065 0.3086 -0.0253* 0.069 Dry season water: river/lake (inverted) 0.1571 0.364 0.1482 0.3556 0.0088 0.602 Toilet: flush 0.4066 0.4914 0.4003 0.4903 0.0064 0.783 Toilet: no facility (inverted) 0.4171 0.4933 0.4097 0.4921 0.0074 0.750 Toilet: shared (inverted) 0.3709 0.4832 0.3693 0.4829 0.0016 0.945 Kitchen type: inside the house (inverted) 0.5061 0.5002 0.4394 0.4966 0.0668*** 0.005 Kitchen type: outside roofed 0.4738 0.4995 0.5404 0.4987 -0.0666*** 0.005 Kitchen type: outside unroofed (inverted) 0.014 0.1174 0.0148 0.1209 -0.0009 0.878 Cooking fuel: wood (inverted) 0.993 0.0833 0.9744 0.1581 0.0186*** 0.003 Cooking fuel: charcoal 0.007 0.0833 0.0256 0.1581 -0.0186*** 0.003 Lighting source: electric network 0.3839 0.4866 0.4623 0.4989 -0.0783*** 0.001 Lighting source: generator (inverted) 0.0271 0.1623 0.0202 0.1408 0.0068 0.332 Lighting source: battery (inverted) 0.0183 0.1342 0.0243 0.154 -0.0059 0.390 Lighting source: kerosene lamp (inverted) 0.062 0.2412 0.0553 0.2286 0.0067 0.543 Lighting source: solar panel (inverted) 0.1894 0.392 0.2102 0.4078 -0.0209 0.270 Nutrition Num. of balls of glutinous rice for women (prior day) 8.8194 8.0422 8.6482 8.1434 0.1711 0.654 Num. of balls of glutinous rice for kids (prior day) 1.4508 3.2564 1.5086 3.3957 -0.0579 0.713 Num. of bowls of ordinary rice for women (prior day, 3.4171 2.8494 3.2884 2.678 0.1287 0.320 inverted) Daily num. of bowls of ordinary rice for kids (prior day, 0.2203 0.8912 0.1157 0.608 0.1046*** 0.002 inverted) Num. of days woman eats meat (weekly) 2.7635 2.2208 2.7102 2.1872 0.0533 0.607 Num. of days woman eats eggs (weekly) 0.5445 1.1778 0.5013 1.0882 0.0432 0.415 Num. of days woman eats dairy (weekly) 0.1091 0.6484 0.0768 0.5125 0.0323 0.230 Num. of days child eats meat (weekly) 0.9223 1.8741 0.746 1.6267 0.1764** 0.030 Num. of days child eats eggs (weekly) 0.2757 0.9997 0.219 0.7868 0.0567 0.170 Num. of days child eats dairy (weekly) 0.0881 0.6897 0.062 0.5705 0.0261 0.371 Kg of meat per HH member (weekly) 0.2385 0.2637 0.2367 0.2939 0.0018 0.892 Kg of fish per HH member (weekly) 0.2021 0.2522 0.2036 0.2487 -0.0015 0.899 Kg of vegetables per HH member (weekly) 0.907 1.0112 0.827 0.8545 0.0801* 0.065 26 Kg of fruits per HH member (weekly) 0.3822 0.5745 0.3181 0.4934 0.0641*** 0.010 Exposure to Shocks Family death shock (inverted) 0.1239 0.3296 0.124 0.3298 -0.0001 0.996 Illness shock (inverted) 0.5646 0.496 0.5593 0.4968 0.0053 0.822 Business or unemployment shock (inverted) 0.0654 0.2474 0.0512 0.2206 0.0142 0.192 Natural disaster shock (inverted) 0.178 0.3827 0.1941 0.3957 -0.0161 0.383 Crop loss shock (inverted) 0.4171 0.4933 0.4461 0.4974 -0.029 0.215 Robbery shock (inverted) 0.6614 0.4734 0.6752 0.4686 -0.0138 0.535 Woman characteristics Woman age (years) 33.0995 10.5313 33.8491 10.2717 -0.7496 0.125 Woman education (years) 2.4485 2.7485 2.6361 3.3249 -0.1876 0.201 Woman is divorced (inverted) 0.0288 0.1673 0.0189 0.1362 0.0099 0.158 Woman is a widow (inverted) 0.0567 0.2314 0.0202 0.1408 0.0365*** 0.000 Woman is married 0.8743 0.3316 0.9205 0.2707 -0.0461*** 0.001 Woman was never married (inverted) 0.0401 0.1964 0.0404 0.1971 -0.0003 0.975 Woman ethnicity: Lao 0.1283 0.3345 0.1186 0.3235 0.0097 0.531 Woman ethnicity: Khmu 0.4319 0.4956 0.372 0.4837 0.0600*** 0.009 Woman ethnicity: Katang (inverted) 0.1134 0.3173 0.1617 0.3684 -0.0483*** 0.003 Woman ethnicity: Other (inverted) 0.3264 0.4691 0.3477 0.4766 -0.0214 0.339 Woman works on HH farm (inverted) 3.5855 2.8454 3.1927 2.8634 0.3928*** 0.004 Woman works in HH business 0.1667 0.9789 0.283 1.3187 -0.1164** 0.039 Woman works in non-HH business 0.9145 2.0791 0.628 1.8519 0.2865*** 0.002 Connections with village head Woman often or sometime had conversations with the 0.7033 1.1038 0.4946 0.9744 0.2087*** 0.000 village head in the past month Someone in HH is a close friend/relative with the village 0.7042 0.4566 0.6644 0.4725 0.0398* 0.071 head * p<0.1 ** p<0.05 *** p<0.01 27 Table A2: Variable Weights for PMT Index PMT Variables Employment and Income Someone in the HH owns a business Someone in the HH is an employee Someone in the HH is self-employed HH has non-labor income (rent, remittances, pension) HH Members 15 to 64 years old None 1 person 2 persons 3 persons 4 persons 5 or more persons HH members 0-14 years old and elderly over 64 years None 1 person 2 persons or more Highest education completed by hh members Vocational school or university Secondary school or lower Housing Conditions Roof material: concrete or wood Wall material: brick or concrete Floor material: marble, ceramic, or tiles Drinking water: pipe water or bottled water Toilet ownership: own toilet Asset Ownership Washing machine Refrigerator Air conditioner TV Mobile phone Telephone Computer Car Motorcycle Bicycle Village Public Services Access to permanent/daily market Access to periodical market at least twice a week Access to lower secondary school Access to upper secondary school Access to road Accessible road during wet season 28 Table A3: Village heads vs PMT – Ranked Sample of 1,245 Households PMT below 64 RMG-eligible Total No Yes No 193 (16%) 294 (24%) 487 (39%) Yes 266 (21%) 492 (40%) 758 (61%) Total 459 (37%) 786 (63%) 1,245 (100%) 29 Table A3: T-Test comparison of village head and PMT targeting methods – Across the full sample Diff. (RMG- Random Diff. (PMT RMG-eligible RMG-eligible Random eligible – PMT poor PMT nonpoor PMT poor PMT nonpoor Sample poor - obs. mean Sample mean Random obs. obs. mean mean obs. nonpoor) Sample) Household composition 1146 742 -0.0151 0.0233 -0.0384** 1187 701 -0.0039 0.0066 -0.0106 Exposure to shocks 1146 742 0.0039 -0.0059 0.0098 1187 701 -0.0047 0.0080 -0.0128 Land and buildings 1146 742 -0.0312 0.0482 -0.0794** 1187 701 -0.0259 0.0438 -0.0696* ownership Nutrition 1146 742 0.0097 -0.0149 0.0246* 1187 701 -0.0423 0.0716 -0.1140*** Livestock ownership 1146 742 -0.0449 0.0693 -0.1142*** 1187 701 -0.0535 0.0906 -0.1442*** Housing conditions 1146 742 -0.0207 0.0320 -0.0527*** 1187 701 -0.0640 0.1084 -0.1725*** Durables ownership 1146 742 -0.0345 0.0533 -0.0879*** 1187 701 -0.0646 0.1093 -0.1739*** Total income per HH 1146 742 -0.0560 0.0865 -0.1426*** 1187 701 -0.3049 0.5164 -0.8213*** member Note: Variables are rendered as composite z-scores, computed by excluding variables used in the creation of the PMT index, and by estimating means and standard deviations for the whole sample of 1,888 observations. * p<0.1 ** p<0.05 *** p<0.01 30 Table A4: T-test comparison of village head and PMT targeting methods – Grouped by village Diff. (RMG- Random Diff. (PMT RMG-eligible RMG-eligible Random eligible – PMT poor PMT nonpoor PMT poor PMT nonpoor Sample poor - obs. mean Sample mean Random obs. obs. mean mean obs. nonpoor) Sample) Household composition 1146 742 -0.0231 0.0357 -0.0589*** 1187 701 0.0149 -0.0252 0.0400*** Exposure to shocks 1146 742 0.0024 -0.0037 0.006 1187 701 -0.0115 0.0195 -0.031 Land and buildings 1146 742 -0.0178 0.0275 -0.0453 1187 701 -0.0213 0.0361 -0.0574 ownership Woman characteristics 1146 742 -0.0134 0.0207 -0.0342*** 1187 701 -0.0079 0.0133 -0.0211* Nutrition 1146 742 0.0061 -0.0095 0.0156 1187 701 -0.0213 0.0360 -0.0573*** Livestock ownership 1146 742 -0.0548 0.0846 -0.1394*** 1187 701 -0.0348 0.0589 -0.0936*** Housing conditions 1146 742 -0.0156 0.0240 -0.0396*** 1187 701 -0.0151 0.0256 -0.0408*** Durables ownership 1146 742 -0.0283 0.0436 -0.0719*** 1187 701 -0.0230 0.0390 -0.0620*** Total income per HH 1146 742 -0.0849 0.1311 -0.2160*** 1187 701 -0.2011 0.3406 -0.5417*** member Note: Variables are rendered as composite z-scores, computed by excluding variables used in the creation of the PMT index, and by estimating means and standard deviations for each village in the sample of 1,888 observations. * p<0.1 ** p<0.05 *** p<0.01 Table A5: T-tests of composite z-cores for RMG selection RMG-eligible Random RMG-eligible Diff. (RMG- P-val obs. Sample obs. mean eligible – 31 Random Random Bonferroni Sample Sample) corrected p- mean val Land and buildings ownership 1146 742 -0.0312 0.0482 -0.0794 0.027 0.213 Household composition 1146 742 -0.0151 0.0233 -0.0384 0.041 0.326 Durables ownership 1146 742 -0.0345 0.0533 -0.0879*** 0.000 0.000 Livestock ownership 1146 742 -0.0449 0.0693 -0.1142*** 0.000 0.001 Total income per HH member (USD PPP) 1146 742 381.919 480.733 -98.8140** 0.004 0.032 Housing conditions 1146 742 -0.0207 0.032 -0.0527*** 0.000 0.002 Nutrition 1146 742 0.0097 -0.0149 0.0246 0.084 0.669 Exposure to shocks 1146 742 0.0039 -0.0059 0.0098 0.666 1.000 RMG-eligible = 1146 observations. Random Sample = 742 observations. All household characteristics are shared as composite z-scores, except for total income per HH member. Variables included are those used in comparisons of RMG and PMT scores. 32 Construction of individual and composite z-scores There are two challenges when analyzing a large number of variables. First, the number of variables and outcomes of t-tests can make it difficult to interpret results. Second, analysis of multiple outcomes increases the risk of Type I error unless the significance tests are adjusted appropriately. Following Kling et al. (2007), we compute composite z-scores as means of individual z-scores for all variables in each group of variables, while inverting variables assumed to be negatively correlated with welfare. We employ the following strategies: • We identify a set of primary outcomes under each of the following groups, noted in Table A1: o Land and buildings ownership; o Household composition; o Durables ownership; o Livestock ownership; o Income; o Housing conditions; o Nutrition; o Exposure to shocks; o Woman characteristics • For income, we use aggregate income per household member since it is better at capturing the underlying notion of welfare than could be achieved through a mean z-score of different types of income. For all other categories of outcomes, we compute composite z-scores through a uniform procedure: o Calculate the z-score of all variables by subtracting the mean and dividing by the standard deviation; o Convert all outcomes so that the signs of all variables in a category go in the same direction of being positively correlated with improved welfare; o In Table A1, all variables negatively correlated with welfare are shown as “inverted,” meaning that their z-scores would be inverted for the composite z-score calculation, though they are not inverted for descriptive statistics or individual z-scores; o Compute composite z-scores for 8 (eight) categories by taking an average of the z-scores of all variables in a given category. • In the computation of composite z-scores used to compare VHT and PMT targeting methods in Figures 6 and 7, and Tables A3 and A4, we exclude variables used in the construction of the PMT index. Hence, the computation procedure is the same as above, but the following variables are excluded when calculating the composite z-scores for their categories: o Land and buildings ownership: • Number of buildings owned (residential and business) o Household characteristics: • HH dependency ratio • Head of household education (years) o Durables ownership: • Num. vehicle (car, van...) • Num. many motorcycle • Num. many bicycle • Num. refrigerator/freezer 33 • Num. washing machine • Num. television • Num. telephone • Num. mobile phone • Num. computer • Num. air conditioner o Housing conditions: • Dwelling walls: brick • Dwelling walls: concrete • Dwelling roof: concrete • Dwelling roof: wood • Dwelling floor: tile • Dry season water: piped water • Dry season water: protected well • Dry season water: unprotected well (inverted) • Dry season water: river/lake (inverted) • Toilet: flush • Toilet: no facility (inverted) • Toilet: shared (inverted) o Woman characteristics: • Woman education (years) • In Figure 7 and Table A4, we compute individual variables’ z-scores by subtracting village-level means and dividing by village-level standard deviations. Composite z-scores are thus averages of village-based z-scores of variables in each category. 34 Importance of vars: Village heads ranking by RMG priority Woman works in non-hh business 1.0000 Kitchen type: outside unroofed 0.7218 Dwelling walls: concrete 0.6918 Num. of days woman eats dairy (weekly) 0.5555 Services income 0.5455 Woman was never married 0.5240 Woman's conversations with village head 0.4975 Dwelling floor: tile 0.4411 Dry season water: unprotected well 0.4381 Head of hh is a widow 0.4305 Lighting source: kerosene lamp 0.4297 Fishes income 0.4256 Lighting source: solar panel 0.4143 Dwelling roof: grass 0.3987 Dwelling walls: brick 0.3832 Dwelling roof: concrete 0.3762 Family death shock 0.3704 Lighting source: battery 0.3699 Business or unemployment shock 0.3615 Head of hh is female 0.3569 Num. of days kids eat eggs (weekly) 0.3565 Num. of days kids eat dairy (weekly) 0.3561 Kitchen type: outside roofed 0.3555 Rainy season water: unprotected well 0.3549 Dry season water: piped water 0.3539 35 Woman is a widow 0.3487 Daily num. of balls of ord. rice for kids 0.3479 Toilet: no facility 0.3446 Rainy season water: piped water 0.3396 Num. of goats 0.3383 Head of hh was never married 0.3344 Dry season water: river/lake 0.3315 Num. of buildings owned 0.3283 Dwelling floor: earth/clay 0.3270 Cooking fuel: wood 0.3268 Natural disaster shock 0.3251 Lighting source: generator 0.3249 Woman ethnicity: Katang 0.3220 Num. of days woman eats eggs (weekly) 0.3201 Woman works in hh business 0.3160 Dwelling roof: tile 0.3147 Salary income 0.3145 Num. of jewelry 0.3139 Head of hh age (years) 0.3116 Toilet: flush 0.3091 Woman ethnicity: Lao 0.3082 Lighting source: electric network 0.3077 Num. of sewing machines 0.3063 Woman ethnicity: Other 0.3052 Num. of days kids eat meat (weekly) 0.3019 Num. of days woman eats meat (weekly) 0.2992 Woman works on hh farm 0.2989 Total income per hh member 0.2977 36 Rainy season water: protected well 0.2958 Head of hh ethnicity: Khmu 0.2935 Toilet: shared 0.2895 Dwelling walls: bamboo 0.2892 Kg of veggies per hh member (weekly) 0.2861 Kg of meat per hh member (weekly) 0.2858 Dwelling floor: wood 0.2857 Crop loss shock 0.2850 Dwelling walls: wood 0.2848 Daily num. of balls of ord. rice for women 0.2842 Cooking fuel: charcoal 0.2839 Illness shock 0.2830 Daily num. of balls of glut. rice for kids 0.2825 Kg of fruits per hh member (weekly) 0.2813 Head of hh ethnicity: Katang 0.2811 Kitchen type: inside the house 0.2795 Extra income 0.2770 Num. of computers 0.2747 Robbery shock 0.2738 Num. of solar panels 0.2732 Kg of fish per hh member (weekly) 0.2731 Dry season water: protected well 0.2722 Num. of telephones 0.2721 Dwelling: num. of rooms 0.2715 Total income 0.2700 Head of hh ethnicity: Lao 0.2684 Handicraft income 0.2680 Rainy season water: river/lake 0.2665 37 Head of hh ethnicity: Other 0.2657 Woman ethnicity: Khmu 0.2648 Head of hh education (years) 0.2643 Dwelling floor: concrete 0.2637 Someone in hh is a friend/relative with village head 0.2596 Agr. income 0.2566 Daily num. of balls of glut. rice for women 0.2551 Head of hh is divorced 0.2549 Woman age (years) 0.2507 Other income 0.2494 Dwelling floor: bamboo 0.2492 Num. of cows 0.2455 Num. of ducks 0.2441 Dwelling roof: metal 0.2410 Woman is divorced 0.2374 Num. of buffalos 0.2320 Num. of washing machines 0.2319 Woman education (years) 0.2285 Forest income 0.2279 Woman is married 0.2251 Num. of TVs 0.2221 Livestock income 0.2205 Num. of satellite discs/connections 0.2158 Num. of radios/vcd 0.2154 Num. of rice mills 0.2134 Num. of food processors 0.2087 HH dependency ratio 0.2085 Num. of chickens 0.2082 38 Head of hh is married 0.2067 Dwelling roof: wood 0.2028 Num. of pigss 0.1946 Num. of mosquito nets 0.1870 Num. of fishing nets 0.1747 Num. of air conditioners 0.1720 Num. of mobile phones 0.1715 Num. of boats 0.1696 Num. of refrigerators/freezers 0.1620 Num. of agr. equipments 0.1391 Num. of two- and four-wheel tractors 0.1359 Num. of motor cycles 0.1354 Num. of electric rice cookers 0.1334 Num. of vehicles (car, vans, etc.) 0.1330 Num. of steam rice cookers 0.1204 Num. of bicycles 0.1049 Size of res. plots (thousand sqm) 0.0852 Size of agr. plots (thousand sqm) 0.0627 Importance of vars: Village heads ranking by land Dwelling walls: bamboo 1.0000 Dwelling: num. of rooms 0.8370 Dwelling walls: concrete 0.8223 Services income 0.8071 Total income 0.7809 Num. of cows 0.7534 Head of hh is a widow 0.7211 Dwelling floor: earth/clay 0.6908 39 Dwelling roof: tile 0.6863 Head of hh age (years) 0.6840 Woman works in hh business 0.6806 Dwelling floor: bamboo 0.6672 Num. of satellite discs/connections 0.6315 Dwelling floor: concrete 0.6170 Head of hh was never married 0.6102 Family death shock 0.6090 Head of hh is female 0.5884 Num. of buffalos 0.5786 Num. of days kids eat eggs (weekly) 0.5781 Dwelling floor: tile 0.5667 Woman ethnicity: Lao 0.5510 Dry season water: unprotected well 0.5469 Num. of mobile phones 0.5369 Lighting source: generator 0.5282 Cooking fuel: wood 0.5274 Rainy season water: river/lake 0.5260 Num. of days kids eat meat (weekly) 0.5000 Total income per hh member 0.4987 Kitchen type: outside roofed 0.4969 Woman is a widow 0.4947 Num. of mosquito nets 0.4938 Daily num. of balls of ord. rice for kids 0.4909 Dry season water: piped water 0.4894 Num. of rice mills 0.4892 Lighting source: solar panel 0.4821 Lighting source: kerosene lamp 0.4805 40 Head of hh ethnicity: Lao 0.4799 Woman was never married 0.4797 Kitchen type: outside unroofed 0.4779 Crop loss shock 0.4775 Woman's conversations with village head 0.4767 Natural disaster shock 0.4739 Num. of days woman eats meat (weekly) 0.4692 Business or unemployment shock 0.4681 Kitchen type: inside the house 0.4593 Dry season water: river/lake 0.4590 Num. of days kids eat dairy (weekly) 0.4579 Num. of TVs 0.4574 Dwelling walls: brick 0.4565 Toilet: flush 0.4551 Dry season water: protected well 0.4531 Lighting source: electric network 0.4514 Rainy season water: piped water 0.4499 Num. of computers 0.4467 Toilet: no facility 0.4458 Woman works in non-hh business 0.4425 Someone in hh is a friend/relative with village head 0.4425 Num. of days woman eats dairy (weekly) 0.4419 Dwelling roof: wood 0.4403 Daily num. of balls of glut. rice for kids 0.4398 Daily num. of balls of ord. rice for women 0.4392 Robbery shock 0.4358 Num. of ducks 0.4339 Kg of fruits per hh member (weekly) 0.4298 41 Rainy season water: protected well 0.4268 Num. of chickens 0.4255 Salary income 0.4244 Livestock income 0.4242 Kg of veggies per hh member (weekly) 0.4238 Toilet: shared 0.4218 Lighting source: battery 0.4188 Dwelling walls: wood 0.4185 Other income 0.4175 Kg of fish per hh member (weekly) 0.4171 Illness shock 0.4150 Agr. income 0.4132 Kg of meat per hh member (weekly) 0.4119 Num. of days woman eats eggs (weekly) 0.4085 Daily num. of balls of glut. rice for women 0.4081 Woman ethnicity: Other 0.4062 Dwelling floor: wood 0.4008 Woman ethnicity: Khmu 0.4004 Num. of buildings owned 0.3969 Head of hh is divorced 0.3962 Woman age (years) 0.3948 Woman education (years) 0.3874 Dwelling roof: grass 0.3811 Handicraft income 0.3792 Woman is divorced 0.3744 Dwelling roof: metal 0.3711 Woman works on hh farm 0.3704 Fishes income 0.3679 42 Head of hh education (years) 0.3659 Extra income 0.3625 Head of hh ethnicity: Katang 0.3610 Head of hh ethnicity: Other 0.3608 Head of hh ethnicity: Khmu 0.3578 Woman is married 0.3560 Rainy season water: unprotected well 0.3546 Num. of goats 0.3424 Woman ethnicity: Katang 0.3386 Head of hh is married 0.3375 Forest income 0.3323 Num. of telephones 0.3304 Cooking fuel: charcoal 0.3215 HH dependency ratio 0.3186 Num. of radios/vcd 0.3164 Num. of jewelry 0.3115 Num. of bicycles 0.3063 Num. of solar panels 0.3060 Num. of pigss 0.3059 Num. of motor cycles 0.2988 Num. of refrigerators/freezers 0.2928 Num. of two- and four-wheel tractors 0.2913 Dwelling roof: concrete 0.2858 Num. of boats 0.2667 Num. of agr. equipments 0.2454 Num. of fishing nets 0.2410 Num. of electric rice cookers 0.2291 Num. of food processors 0.2195 43 Num. of sewing machines 0.2171 Num. of vehicles (car, vans, etc.) 0.1867 Size of agr. plots (thousand sqm) 0.1704 Num. of steam rice cookers 0.1670 Size of res. plots (thousand sqm) 0.1662 Num. of washing machines 0.1430 Num. of air conditioners 0.0760 Importance of vars: Village heads ranking by hunger Dwelling walls: concrete 1.0000 Dwelling floor: tile 0.9515 Dwelling: num. of rooms 0.7716 Num. of motor cycles 0.7146 Total income 0.6791 Dwelling walls: bamboo 0.6732 Woman works in hh business 0.6621 Head of hh is female 0.6372 Num. of buffalos 0.6142 Dwelling roof: tile 0.6003 Head of hh ethnicity: Katang 0.5282 Num. of mobile phones 0.5218 Dwelling floor: concrete 0.5099 Rainy season water: unprotected well 0.5055 Services income 0.4988 Toilet: no facility 0.4792 Head of hh is a widow 0.4777 Woman is a widow 0.4772 Dwelling floor: earth/clay 0.4763 44 Dwelling roof: grass 0.4707 Total income per hh member 0.4702 Lighting source: kerosene lamp 0.4684 Dry season water: unprotected well 0.4680 Num. of refrigerators/freezers 0.4678 Kitchen type: inside the house 0.4620 Daily num. of balls of ord. rice for kids 0.4541 Num. of cows 0.4441 Num. of buildings owned 0.4438 Kitchen type: outside unroofed 0.4417 Kitchen type: outside roofed 0.4413 Livestock income 0.4412 Lighting source: battery 0.4354 Business or unemployment shock 0.4239 Dwelling floor: bamboo 0.4231 Num. of TVs 0.4214 Dry season water: protected well 0.4199 Toilet: flush 0.3972 Fishes income 0.3971 Dwelling walls: brick 0.3963 Lighting source: generator 0.3942 Lighting source: solar panel 0.3939 Family death shock 0.3935 Num. of days kids eat eggs (weekly) 0.3928 Rainy season water: protected well 0.3914 Daily num. of balls of ord. rice for women 0.3897 Num. of days woman eats meat (weekly) 0.3851 Lighting source: electric network 0.3845 45 Toilet: shared 0.3839 Num. of days woman eats dairy (weekly) 0.3827 Rainy season water: piped water 0.3806 Woman is divorced 0.3798 Woman education (years) 0.3782 Woman ethnicity: Lao 0.3778 Robbery shock 0.3751 Num. of rice mills 0.3723 Someone in hh is a friend/relative with village head 0.3718 Kg of fruits per hh member (weekly) 0.3683 Woman's conversations with village head 0.3681 Woman was never married 0.3668 Illness shock 0.3646 Natural disaster shock 0.3610 Dry season water: piped water 0.3609 Head of hh ethnicity: Lao 0.3571 Head of hh is divorced 0.3557 Num. of satellite discs/connections 0.3549 Kg of veggies per hh member (weekly) 0.3534 Woman ethnicity: Katang 0.3533 Dry season water: river/lake 0.3532 Num. of days kids eat meat (weekly) 0.3528 Crop loss shock 0.3519 Num. of chickens 0.3497 Cooking fuel: wood 0.3484 Rainy season water: river/lake 0.3472 Woman works on hh farm 0.3467 Kg of fish per hh member (weekly) 0.3460 46 Daily num. of balls of glut. rice for women 0.3447 Num. of mosquito nets 0.3439 Woman ethnicity: Other 0.3405 Kg of meat per hh member (weekly) 0.3403 Dwelling floor: wood 0.3393 Num. of days woman eats eggs (weekly) 0.3379 Daily num. of balls of glut. rice for kids 0.3365 Head of hh education (years) 0.3319 Head of hh ethnicity: Khmu 0.3296 Dwelling walls: wood 0.3290 Woman ethnicity: Khmu 0.3268 Woman works in non-hh business 0.3240 Agr. income 0.3203 Head of hh was never married 0.3201 Handicraft income 0.3183 Num. of sewing machines 0.3109 Extra income 0.3071 Head of hh ethnicity: Other 0.3053 Num. of jewelry 0.2993 Salary income 0.2979 Woman is married 0.2957 Woman age (years) 0.2932 Head of hh is married 0.2905 Other income 0.2899 Head of hh age (years) 0.2878 Dwelling roof: metal 0.2846 Num. of pigss 0.2796 Num. of goats 0.2701 47 Forest income 0.2689 Num. of ducks 0.2685 Num. of days kids eat dairy (weekly) 0.2653 Num. of solar panels 0.2541 HH dependency ratio 0.2532 Cooking fuel: charcoal 0.2510 Num. of telephones 0.2484 Num. of vehicles (car, vans, etc.) 0.2388 Num. of boats 0.2347 Num. of radios/vcd 0.2324 Num. of fishing nets 0.2307 Num. of two- and four-wheel tractors 0.2216 Dwelling roof: wood 0.2153 Num. of bicycles 0.2066 Num. of agr. equipments 0.2030 Num. of food processors 0.1928 Num. of electric rice cookers 0.1919 Num. of computers 0.1732 Dwelling roof: concrete 0.1640 Num. of steam rice cookers 0.1598 Size of agr. plots (thousand sqm) 0.1547 Size of res. plots (thousand sqm) 0.1530 Num. of washing machines 0.1488 Num. of air conditioners 0.1318 Importance of vars: Village heads ranking by need Dwelling: num. of rooms 1.0000 Services income 0.9838 48 Head of hh is female 0.9774 Num. of mobile phones 0.8470 Woman works in hh business 0.7672 Dwelling walls: concrete 0.7033 Dwelling walls: bamboo 0.6860 Total income 0.6722 Woman is a widow 0.6459 Dwelling roof: tile 0.6241 Head of hh is a widow 0.6015 Lighting source: kerosene lamp 0.5926 Dry season water: unprotected well 0.5660 Num. of motor cycles 0.5571 Dwelling floor: tile 0.5508 Num. of days kids eat meat (weekly) 0.5202 Num. of days kids eat dairy (weekly) 0.5187 Kitchen type: inside the house 0.5035 Livestock income 0.4971 Dwelling roof: grass 0.4965 Daily num. of balls of ord. rice for kids 0.4933 Toilet: no facility 0.4900 Kitchen type: outside roofed 0.4886 Num. of buildings owned 0.4823 Num. of buffalos 0.4807 Lighting source: generator 0.4652 Business or unemployment shock 0.4580 Num. of cows 0.4559 Total income per hh member 0.4530 Lighting source: battery 0.4505 49 Lighting source: solar panel 0.4494 Dwelling floor: concrete 0.4437 Kg of fish per hh member (weekly) 0.4431 Num. of days kids eat eggs (weekly) 0.4424 Dwelling floor: bamboo 0.4421 Kitchen type: outside unroofed 0.4418 Cooking fuel: wood 0.4339 Dwelling floor: earth/clay 0.4311 Num. of rice mills 0.4292 Family death shock 0.4283 Woman works in non-hh business 0.4189 Woman was never married 0.4182 Num. of refrigerators/freezers 0.4153 Dry season water: piped water 0.4148 Head of hh ethnicity: Katang 0.4130 Dry season water: protected well 0.4114 Toilet: flush 0.4083 Toilet: shared 0.4066 Head of hh is divorced 0.4013 Woman ethnicity: Other 0.3978 Fishes income 0.3969 Daily num. of balls of glut. rice for kids 0.3958 Head of hh ethnicity: Lao 0.3876 Rainy season water: piped water 0.3874 Head of hh ethnicity: Other 0.3872 Rainy season water: unprotected well 0.3858 Natural disaster shock 0.3817 Kg of veggies per hh member (weekly) 0.3811 50 Salary income 0.3805 Woman's conversations with village head 0.3778 Num. of satellite discs/connections 0.3775 Woman ethnicity: Lao 0.3744 Num. of days woman eats dairy (weekly) 0.3729 Num. of days woman eats eggs (weekly) 0.3711 Rainy season water: river/lake 0.3710 Kg of meat per hh member (weekly) 0.3702 Num. of days woman eats meat (weekly) 0.3698 Kg of fruits per hh member (weekly) 0.3666 Illness shock 0.3645 Robbery shock 0.3643 Daily num. of balls of glut. rice for women 0.3632 Extra income 0.3627 Someone in hh is a friend/relative with village head 0.3624 Other income 0.3618 Dwelling floor: wood 0.3608 Rainy season water: protected well 0.3591 Dry season water: river/lake 0.3570 Woman is divorced 0.3558 Daily num. of balls of ord. rice for women 0.3556 Dwelling walls: brick 0.3521 Lighting source: electric network 0.3510 Dwelling roof: wood 0.3510 Crop loss shock 0.3483 Dwelling walls: wood 0.3474 Num. of TVs 0.3452 Num. of mosquito nets 0.3441 51 Num. of chickens 0.3389 Head of hh ethnicity: Khmu 0.3378 Woman ethnicity: Katang 0.3358 Woman education (years) 0.3292 Agr. income 0.3260 Dwelling roof: metal 0.3256 Num. of ducks 0.3252 Num. of pigss 0.3247 Woman works on hh farm 0.3234 Handicraft income 0.3220 Woman age (years) 0.3167 Head of hh education (years) 0.3143 Num. of goats 0.3099 Head of hh age (years) 0.3098 Woman ethnicity: Khmu 0.3060 Forest income 0.3047 Head of hh is married 0.3014 Woman is married 0.2927 HH dependency ratio 0.2851 Num. of sewing machines 0.2806 Cooking fuel: charcoal 0.2797 Num. of telephones 0.2779 Num. of air conditioners 0.2762 Num. of jewelry 0.2650 Num. of solar panels 0.2641 Num. of vehicles (car, vans, etc.) 0.2447 Num. of two- and four-wheel tractors 0.2382 Num. of radios/vcd 0.2298 52 Num. of electric rice cookers 0.2189 Head of hh was never married 0.2178 Num. of boats 0.2130 Num. of fishing nets 0.1945 Num. of agr. equipments 0.1837 Num. of food processors 0.1681 Num. of steam rice cookers 0.1552 Num. of washing machines 0.1524 Num. of computers 0.1514 Dwelling roof: concrete 0.1501 Num. of bicycles 0.1399 Size of agr. plots (thousand sqm) 0.1173 Size of res. plots (thousand sqm) 0.1170 53