Should I Join More Homogenous or Heterogeneous Social Networks? Empirical Evidence from Iddir Networks in Ethiopia Kibrom A. Abay Department of Economics University of Copenhagen Email: Kibrom.Araya.Abay@econ.ku.dk Guush Berhane Development Strategies and Governance Division International Food Policy Research Institute Email: Guush.Berhane@cgiar.org January 16, 2016 Abstract This paper empirically characterizes alternative compositional formation of social networks and their implications on various costs and benefits to members of the network. We investigate the implication of compositional differences in social networks, whether majority of members are from the same village, using data from iddir networks in Ethiopia. We exploit households subscription to more than one iddir networks and rely on fixed-effect approaches to compare with-in household variation in costs and benefits associated with different iddir networks. We empirically show that compositionally homogenous iddir networks are potentially easier to monitor and influence, while more heterogeneous networks are more likely to suffer from these challenges. Furthermore, we document that more heterogeneous iddir networks are more likely to marginalize households with low social status or low membership record. Interestingly, we finally find some evidence which suggest that more homogenous iddir networks (those dominated by members of the same village) are more effective in dealing with idiosyncratic shocks that affect few members in the village while less effective in insuring covariate shocks that may affect many members in the group. Keywords: Social networks, iddir networks, homogenous and heterogeneous networks, compositional differences in social networks, heterogeneous access and benefits. Corresponding author. This paper benefited from comments by Bruce Wydick and Ulrik Richardt Beck. All remaining errors are ours.
1. Introduction Modern economics literature acknowledges that social networks play crucial role in shaping individuals behavior and economic outcomes. Social networks play important role in facilitating technology adoption (Foster and Rosenzweig, 1995; Conley and Udry, 2002; Bandiera and Rasul, 2006), mitigating market inefficiencies (Udry, 1990; de Janvry et al., 1991; Wydic, 1999; Abay et al., 2014), and providing risk-sharing mechanisms (Fafchamps and Lund, 2003; Wydick et al., 2011; Dercon et al., 2012; Kinnan and Townsend, 2012). 1 On the other hand, some other studies show that they can lead to resource misallocation (Banerjee and Munshi, 2004); marginalize outsiders in a society (Agarwal, 2001); create uneven playing field in the labor market (Granovetter, 1995). These contradictions necessitate further theoretical and empirical research to identify which type and when social networks can provide better benefits to members and society at large. While there is some evolving theoretical literature that justifies which types of social networks are efficient (Jackson and Wolinksky, 1996; Bala and Goyal, 2000; Jackson and van den Nouweland, 2005), the empirical strand of this literature is not conclusive. More specifically, the following question deserve further empirical investigation: What is the optimal composition of social networks? Empirical characterization and identification of the potential of social networks is often prone to a lair of empirical problems. First, and as argued in Manski (1993), endogenous formation of social networks challenges identifying the gains and benefits from social networks and social interactions. Secondly, some qualities of social networks are more suited for pursuing a specific individual and societal objective than others. For instance, some previous studies suggest that heterogeneous social ties are more likely to generate economic opportunities than more clustered and homogenous networks (Newman, 2003; Page, 2007; Eagle et al., 2010). Along this line, Krishnan and Sciubba (2009) characterize labor-sharing arrangements in Ethiopia and show that asymmetric arrangements (those among farmers of heterogeneous quality) have positive implications on economic performance. However, from social learning and efficient monitoring perspective, more homogenous networks facilitate smooth flow of information (Ellison and Fudenburg, 1993; Munshi, 2004) and effective monitoring among group members (Huppi and Feder, 1990; Wydick, 1999; Karlan, 2007). More homogenous social networks are also relatively less susceptible to free-rider problem (Granovetter, 2005), and more suitable (and attractive) for 1 Durlauf and Fafchamps (2005), and Jackson (2008) provide excellent review on the formation and relevance of social networks. 1
informal risk-sharing (Delpierre et al., 2014). 2 Thirdly, the aforementioned properties of social networks may encourage strategic choice of network types for different purposes. For instance, some social networks are potentially more effective in dealing with idiosyncratic shocks that affect few members of the network while some other networks may serve as insurance institutions for dealing covariate shocks that affect a good portion of members of their group. Thus, if a household subscribes to a more homogenous social network for social interaction (socializing) purpose while also subscribing to other heterogeneous network for insurance purposes, we cannot attribute the gains and benefits from these different networks as qualities of these networks. In this paper, we study an indigenous social network in Ethiopia, namely iddir associations. Iddir is the most widespread social network in Ethiopia, commonly established by community members, neighbors, or among friends and families, for serving well-defined objectives. The main objective of iddir associations is to provide funeral services and to support bereaved family members morally and financially (Dercon et al., 2006; Abay et al., 2014). While iddir associations are well-established social networks in Ethiopia, very little is known about how and which types of iddir networks are effective in serving their members and society at large. In this study, we empirically characterize alternative formation of iddir networks and their implications in terms of the various costs and benefits for members. We specifically focus on the implication of compositional differences in iddir networks, whether majority of iddir members are from the same village or not. We first document that iddir networks dominated by members from the same village are compositionally more homogenous than those networks that cover beyond a village. 3 We then characterize these potentially homogenous and heterogeneous iddir networks considering various costs and benefits to members as well as the potential of these different types of social networks in insuring different types of shocks. Such characterization helps to identify the potential of alternative social networks in facilitating insurance and credit facilities in developing countries. We are not the first to study the implication of compositional structure in social networks. A recent theoretical study by Delpierre et al. (2014) show that group composition of informal risksharing networks have varying implications on the insurance coverage of poor and rich household members of the group. 2 Delpierre et al. (2014) investigate the implication of compositional structure in informal risk-sharing arrangements and provide some theoretical predictions related with insurance coverage of different members of the group. In particular, poor and rich households may have varying level of insurance coverage from risk-sharing networks with different composition. 3 This is commonly referred as homophily in the sociological literature (McPherson et al., 2003; Newman, 2003; Golub and Jackson, 2012). 2
We address the problem of endogenous formation of social networks and strategic choice of network types by exploiting households subscription to more than one type of social networks and hence rely on fixed-effect approaches. Using a detailed household-level data that provides within household variation in the choice of iddir type, we compare with-in household variation in the benefits and costs that a household gains from subscribing to two (or more) compositionally different iddir networks. We augment this comparison using detail observational information about households motives (and objectives) for joining each iddir network they subscribe. By doing so, we cancel-out individual (household) level effects that affect iddir type choice while also controlling for households strategic network choice for different purposes driven by some welldefined motives. We also consider a different sampling strategy that provides more exogenous network formation (type choice) by considering only those households who inherited their iddir networks from their parents. We empirically show that compositionally homogenous iddir networks are potentially easier to monitor as they are able to mobilize in-kind resource, while more heterogeneous networks are more likely to suffer from this challenge. We find that households subscribing to iddir networks dominated by members from the same village have slightly higher access to loan service from their iddir networks. This has some implications in terms of monitoring free-rider problem in collective actions involving various compositional structure and this is consistent with the existing credit literature which justify that peer monitoring is easier in more homogenous groups (see, Huppi and Feder, 1990; Wydick, 1999; Karlan, 2007). We also find that households joining iddir networks dominated by their type have higher probability of influencing their networks. Furthermore, we document heterogeneous access and benefits among households subscribing to compositionally different iddir types. Most importantly, we find that more heterogeneous iddir networks are more likely to marginalize households with low social status or low membership record. These results show that marginalized households, those with short record of iddir membership or those with low social status, are better off by joining more homogenous iddir networks, those iddir networks composed of their type. This has some implications in terms of designing policy interventions to support social networks that can serve marginalized and poorer rural households. We finally find some evidence which highlights iddir networks dominated by members of the same village are more effective in dealing with idiosyncratic shocks that affect few members in the village while less effective in insuring covariate shocks that may affect many members in the group. 3
2. Iddir Networks in Ethiopia Iddir is the most pervasive type of social network in Ethiopia, which operates both in rural and urban settings. Iddir networks are funeral associations meant to deal with material and moral costs associated with death of members and their families. They provide payouts, and other material and moral support when a family member dies while also assuming a key role in facilitating the funeral of the deceased member. Payouts as well as other types of support (in-kind, for instance, food and labor) are generated from members as monthly contributions or at the time of funeral. While iddir networks are meant to serve as funeral associations, they go beyond that and are involved in serving some other defined objectives. 4 They often provide small-scale credit for their members (Dercon et al., 2006), and assist members suffering from other types of shocks (Pankhurst and Mariam, 2000; Mariam, 2003; Dercon et al., 2006). Iddir networks own interesting features that liken them with other semi-formal risk-sharing social networks that operate under weak regulatory enforcements. They offer informal social insurance, serve as information hub for members, and help strengthen trust among members, microlevel properties that other types of social networks share (Caeyers and Dercon, 2012). Although iddirs operate without any regulatory environment and systematic incentive for enforcing agreements, they own a strong social leverage that iddirs and their leaders are bestowed from members. These include group pressure and social penalties on individuals that fail to comply with agreed terms. This social pressure and social stigma on members failing to comply with iddir rules is a potential force that keeps these networks sustain long. Most iddirs operate within village and are commonly open to all members living in the village. They have well-defined and sometimes written rules associated with membership contributions, criteria for making payouts and beyond (Dercon et al., 2006). While there might be some variation on how iddirs operate, particularly rural-urban variation, iddir networks in Ethiopia share some fundamental characteristics that are commonly practiced in many parts of Ethiopia. Participation in iddir networks is pervasively common in many parts of Ethiopia. Bold and Dercon (2014) report that around 97% of households in their sample belong to one or more iddir networks. This is slightly lower in our sample, where around 90% of the households in our sample are members of one or more iddirs. Members are required to pay monthly contributions and some occasional contributions at the time of funeral or in times when the iddir calls for additional 4 This is unlike many other social networks that are formed for serving generic and unlimited interest of members. 4
contributions. Membership fees vary across iddirs and are higher in urban areas than rural Ethiopia. For instance, Dercon et al. (2006) report that the average monthly household contribution (to iddir) in their sample amounts 1.64 Ethiopian Birr, while this amounts around 3.4 Ethiopian Birr (around 0.17 USD) in our data. In terms of size, iddir networks comprise varying number of households. Dercon et al. (2006) report that an average iddir in their sample comprises about 84 members (with a median of 55), while the corresponding number in our sample amounts 78 households (with a median of 55 households). Although iddirs are mainly formed to support funeral services to members suffering death of a family member, they often provide other services to members including small-scale loans and social support in times of hardship because of shocks. Almost all iddirs provide commonly agreed payouts (in cash) when members suffer death of a family member. Dercon et al. (2006) report that, on average, iddir associations in their sample pay around 206 Ethiopian Birr as payout, while the corresponding rate in our sample amounts 537 Birr. 5 Despite few recent attempts to investigate the potential of iddir networks in Ethiopia, more rigorous analysis on which type of iddirs are more beneficial and effective is not yet available. 6 In particular, more rigorous characterization of the micro-level properties of these networks and their implication in terms of the gains and costs to members is missing. 7 This is not surprising, given that similar empirical characterization of social networks at broader perspective is scant. We aim to fill this gap by characterizing two types of iddir associations in Ethiopia, those dominated by members of the same village and those who are not. We first show that those iddir associations dominated by members from the same village are more homogenous than those associations that go beyond a village. We then empirically characterize these two types of iddir networks in view of: (a) the motives and objectives they are thought to provide; (b) the costs, benefits and services they provide to members; (c) the potential of members to influence their networks and overall implication on monitoring their network; (d) the potential of these different types of networks in insuring different types of shocks. 5 To put this amount in some perspective in terms of consumption per household member in our sample, the average payout amounts around 3 months average consumption per member in our sample. The average weekly consumption per member in our sample is around 42 Birr. Currently, 1 USD 20 Ethiopian Birr. 6 Previous studies have mainly investigated potential determinants of iddir participation. Dercon et al. (2006) show that demographic attributes of households including age and household size affect iddir participation, while socioeconomics measures such as wealth, land, and livestock holdings had no effect (see also, Hoddinott et al., 2005; Abay, 2014). 7 Bold and Dercon (2014) might be an exception to this, as they compare the performance of iddir associations that operate with considerable savings and those without savings, in terms of insuring their members. 5
3. Data Source and Sampling Design The data we use here comes from a randomized control trial conducted for evaluating the demand for weather-index insurance and the role of informal risk-sharing networks, particularly iddir, in Ethiopia. The data is collected by the International Food Policy Research Institute (IFPRI) and the University of Oxford. The randomization is made at both household and iddir level and covers four woredas (districts) in Oromia region of Ethiopia (see Berhane et al., 2013 for details on the sampling design). The sites covered include Adama, Shashemene, Dodota, and Bako-Tibe. From these four sites an equal number of villages (totally 110) are randomly selected and from which a total of around 2400 households are selected randomly. These four sites are located close to each other. However, since the randomizations were made for a different purpose, it only provides us a random sample of households and iddirs, and not random assignment of households to different iddir types. The first survey (baseline survey) was conducted in February-March 2011, and three follow-up surveys were added at latter years. But the information about household participation in iddirs is more thorough and complete in the baseline survey (first survey). For this reason, we will mainly use the first (baseline) survey in this study while also using the latter surveys to generate some outcomes of interest. Since the randomization was made at household and iddir level we have a random of households and iddirs operating in the study area. While the household level data provides rich information on iddir types for each iddir the household subscribes, the iddir level data also augments this with some aggregate statistics. However, merging these two data sets is not straightforward and leads to substantial reduction in sample size, because both randomizations are made independently and hence the randomly selected iddirs do not cover those listed by households. This is further complicated by some inconsistencies in spelling iddir names in both data sets, and for these reasons we will heavily rely on the household-level data while also using the aggregate iddir level data to supplement the main household-level analysis. Note that as many households in our sample are members of more than one iddir network, our data set still assumes some panel structure. Table 1 provides summary statistics of the characteristics of households in our sample. The table shows that around 85% of the sampled households are male-headed, and the average family size is 6 while the average land size amounts 1.3 hectares. The average age of household head is 43 years and most households are illiterate, with an average education in the sample of 2.4 years of schooling. Households in our sample have good stock of social capital. On average households in our sample can call and rely on 6 individuals in time of need. Most households in our sample 6
participate in other local networks including iddir, equib and cooperatives operating in their villages. 8 On average a household subscribes to two iddirs (with a median subscription of two), and close to 90% of the sample report that they are members of one or more iddir networks. Table 1 also shows that around 7% of households have received transfers from their iddir networks in the last 12 months. Our survey also elicits deep behavioral characteristics of households including subjective measures of risk-taking behavior, trust levels on various agents and level of liquidity constraint. Table 1: Summary Statistics of Sampled Households Explanatory variables Observations Mean Standard deviation Gender of the household head (1=male) 2397 0.847 0.360 Age of household head 2394 42.724 14.471 Education of household head 2386 2.426 2.809 Household size 2398 5.905 2.337 Household head born in this village 2397 0.819 0.385 Household head hold position in Kebele 2392 0.262 0.439 Land size owned 2398 1.331 1.414 Value of livestock asset owned (Birr) 2398 8389.845 9434.189 Number of oxen owned 2398 1.125 1.159 Self-reported wealth status 2391 3.542 1.095 Household owns mobile phone 2398 0.335 0.472 Travel time to the nearest market (minutes) 2341 169.025 154.950 Social capital (people you can rely on in time of need) 2395 5.590 14.128 Household is member of iddir 2398 0.891 0.312 Number of iddirs the household subscribes 2398 2.002 1.190 Household is member of equib 2386 0.086 0.281 Household is member of cooperative 2393 0.275 0.447 Father of household respected in village 2398 0.728 0.445 Mother of household respected in village 2398 0.574 0.495 Parents hold position in Kebele 2392 0.173 0.378 Relatives hold position in Kebele 2392 0.495 0.500 Religion: Muslim 2398 0.485 0.500 Religion: Orthodox 2398 0.286 0.452 Religion: Protestant 2398 0.209 0.407 Liquidity constraint 2397 1.113 0.848 Generalized trust level 2396 2.383 0.789 Trust level on iddir leaders 2363 2.892 0.691 8 Equib is a form of rotating credit and saving association in Ethiopia, which commonly functions as a source of informal finance for member through which each member periodically contributes small sum to get one large sum of it at some point. Equib and iddir associations have distinct features and objectives as the former mainly function as a financial intermediary, rather than as an inclusive social network of broader purpose. 7
Risk taking behavior (level) 2389 3.200 1.520 Household received transfer from iddir in the last 12 months 2398 0.067 0.249 Notes: This table provides descriptive statistics of the explanatory variables considered in the analysis measured at household level. The first column presents the number of observations, while the second and third columns provide mean values and standard deviations, respectively. Kebele is the smallest administrative unit in Ethiopia and it corresponds to peasant association or village. Our empirical focus in this paper is on those households who are iddir members (close to 90% of the sample in Table 1). In Table 2 we provide detail characteristics of each iddir the household subscribes. On average households remained iddir members for around 21 years. Around 79% of the iddir associations in our sample are dominated by members from the same village. This is our variable of interest and we aim to characterize the implication of joining social networks that very much likens one s own characteristics, compared to joining social networks outside your vicinity as stranger. The same table shows that 34% of the iddir subscriptions are inherited from parents. This is also interesting information that provides more exogenous choice of network type, and we are going to exploit this information in some of our robustness exercises in Section 7. Households are asked about the motive for joining each iddir network they subscribe, beyond the common funeral support that all iddirs provide. This information may capture households different and strategic motives for joining a specific type of iddir network. In around 5% of the cases households aim for getting loan; in 21% of the cases they aim for getting protection (insurance) from the network; while the major motive (beyond funeral support) appears to be for social interaction (socializing) purpose (43%). The average monthly contributions in these iddir associations amounts around 3.4 Birr, while average contributions at the time of funeral amounts 2 Birr. Furthermore, members are required to contribute in-kind at funerals, which on average is valued to be around 6 Birr in our sample. These in-kind contributions commonly include food, material support and labor services at the time of funeral. Table 2 also shows that around 44% of iddirs provide loans to members. Admittedly, we do not have information on the exact size of each iddir network in the household-level survey while we have it in the iddir level survey (see, the iddir-level analysis in Section 7.2). As a solution to this, we generate a technically valid proxy for the size of iddirs based on the frequency of iddir names mentioned in our sample. Given that we randomly select households independent of their iddir membership (in each village), the frequency of iddir names mentioned can serve as valid proxy for the size and coverage of iddirs in the study area. 9 Furthermore, we later formally show (in 9 However, some unsystematic spelling errors in naming of iddirs may not be ruled out. Given that these errors are not potentially systematic; the consequence of this measurement error can be predicted. 8
Section 7.2) that the true size of iddirs is not related with our variable of interest, composition of iddirs using the aggregate iddir-level data. Table 2 further shows that in around 13% of the cases households are office bearers of their iddir(s); 16% have received payouts from their iddir in the last three years; and 7% of households have been denied a service they deserve from their iddir network. These figures measure the influence of households on their iddir networks and may hint potential differential treatment of iddir networks for different households. Investigating and attributing these pieces of descriptive evidence to the various characteristics of iddir networks is crucial to understand which type of social networks are more beneficial to members and society at large. Table 2: Households Iddir Membership and their Characteristics Explanatory variables Observations Mean Standard deviation Years since the household is a member of this iddir 4739 20.510 9.470 The husband is the member of this iddir 4795 0.650 0.477 The wife is the member of this iddir 4795 0.321 0.467 Other members of the household are members 4795 0.029 0.168 Majority of members are from this (same) village 4766 0.788 0.409 Parents used to be members of this iddir 4674 0.335 0.472 Motive for joining this iddir (beyond funeral support): Getting loan 4795 0.050 0.218 Getting protection 4795 0.211 0.408 For socialization purpose 4795 0.425 0.494 For making friends 4795 0.070 0.255 Other motives 4795 0.206 0.404 Monthly cash contributions (Birr) 4756 3.433 3.589 Cash payments at the time of funeral (Birr) 4610 2.055 4.874 In-kind contribution at the time of funeral (Birr) 4548 5.983 11.967 Iddir has regular meeting 4795 0.707 0.455 Frequency of these meetings (annually) 3388 5.061 5.245 Does the iddir provide loans 4795 0.435 0.496 Iddir had request for additional contributions in the last three years 4795 0.349 0.477 Are you office-bearer of this iddir? 4795 0.127 0.333 Iddir gives gifts for events other than funeral (e.g., wedding) 4795 0.302 0.459 Circumstances for giving payouts: In case fire destruction 4795 0.184 0.388 Payout for wedding 4795 0.030 0.171 Payout in case of illness 4795 0.035 0.185 Proxy for size of iddir 4795 6.362 15.042 Has the household received payouts in the last three years 4795 0.161 0.368 Have you been denied a service that the household deserve 4795 0.071 0.257 9
Notes: This table provides descriptive statistics of iddirs and households measured for each iddir the household subscribes. The first column presents the number of observations, while the second and third columns provide mean values and standard deviations, respectively. We supplement the above household-iddir-level characterization using an iddir-level descriptive statistics on a random sample of 138 iddir networks operating in the study area. Table 3 provides descriptive figures from these 138 iddir networks. The table shows that the average size of iddir networks in the area is 78 members (with a median of 55), and around half of these members are from the same village. Note that the latter figure picks slightly distinct information than the compositional differences we are exploiting in the household-level data, whether the majority of iddir members are from the same village or not. On average iddir networks in the sample area existed for more than 25 years. This shows that iddirs have sustained long as robust community networks. Most of the iddir networks (99%) have regular contributions and around 15% of these iddirs have bank account. The average payout for a deceased member amounts 537 Birr, a figure slightly higher than the amount reported in Dercon et al. (2006), and Bold and Dercon (2014). Table 3: Aggregate Level Statistics from the Iddir Level Data Explanatory variables Observations Mean Standard deviation Years since the iddir was formed 137 25.073 14.360 Size of the iddir 138 78.261 66.633 Members from this (the same) village 138 37.203 53.522 Iddir require regular contribution 138 0.986 0.120 Size of regular contributions 136 5.487 10.400 Iddir has bank account 138 0.145 0.353 Size of payout for a deceased (family) member 138 537.210 502.774 Notes: This table provides aggregate summary statistics for a random sample of 138 iddirs operating in the sampled woredas (districts). The first column presents the number of observations, while the second and third columns provide mean values and standard deviations, respectively. 3.1 Are iddir networks dominated by members of the same village more homogenous? Before embarking on the main characterization of network formation, it is crucial to empirically show that iddir networks dominated by members of the same village are compositionally more homogenous (in terms of households characteristics) than those iddir networks that cover beyond a village. To do so, we employ self-reported iddir names provided by households and registered by enumerators. Using these iddir names we compute some commonly used empirical measures of compositional heterogeneity (see, Krishnan and Sciubba, 2009 for a similar exercise). We focus on standard deviations of household characteristics among iddir members. For each of these household 10
characteristics, we conduct a simple t-test of whether iddir networks dominated by members from the same village are more homogeneous than those iddir networks that cover beyond a village. As we suspect the accuracy of spelling of iddir names and also to avoid size effects, we consider alternative criteria to compare the level of heterogeneity among iddir members of different type by restricting the frequency of iddir names (proxy for iddir size) at various levels. Column 1 and 2 of Table A1 (in appendix) provide standard deviation estimates for each household characteristic by restricting the frequency of iddir names to be below the median iddir size in the study area (55 members). Columns 3 and 4 report standard deviations by restricting the frequency of iddir names to be less than 10, while columns 5 and 6 restrict the frequency of iddir names to be less than 5. The standard deviations for the various characteristics of households in Table A1 show that iddir networks dominated by members from the same village are characterized by lower compositional heterogeneity. In all exercises, the standard deviation values (across members of the same iddir) for many household characteristics are statistically lower for those iddir networks dominated by members from the same village. In all comparisons, we are not able to observe any case where standard deviation value for a variable is larger for those households in iddir networks dominated by members of the same village. This is theoretically expected because households living in the same village are expected to be observationally similar compared to those households living in different villages. Our results show that the compositional heterogeneities between the two types of iddirs are particularly significant in terms of households educational status, self-reported economic status, and social capital endowment. Interestingly, the differences in standard deviations are more visible and significant when we restrict the size (frequency of iddir names) to be smaller. As we move from columns 1-2, the differences in standard deviations in columns 3-4 and 5-6 almost double and these differences are consistently signed. These descriptive results strongly support our theoretical motivation that justifies iddir networks dominated by members from the same village are more homogenous than those that cover beyond a village. These descriptive results are also consistent with the empirical and theoretical predictions by Krishnan and Sciubba (2009). 10 10 In a slightly similar characterization, Krishnan and Sciubba (2009) show that the number of blood relatives in a village (or network) can predict type of network formation. They particularly show that those rural farmers with large number of close blood relatives in a village are more likely to be engaged in symmetric (homogenous) networks. 11
4. Empirical Characterization of Iddir Networks 4.1 Econometric Method Empirical characterization of social networks is challenging for several reasons. A well-documented challenge is endogenous formation of social networks (Manski, 1993). There are two lairs of endogeneity problems in this regard. The first and the outer level is endogenous formation or selfselection of individuals into a social network. A second and inner lair of this decision involves a choice on which type of networks to join. Since we are interested in characterizing the qualities of alternative formation of social networks, the first problem is not a concern in our case. Thus, we focus on addressing the second problem. There are two ways to address the problem of endogenous selection (sorting) of network type: (a) a straightforward (and first-best) solution is random assignment of individuals into different type of networks. (b) In the absence of quasi-experimental data that assigns households into different network types exogenously, we may rely on fixed-effect approaches if individuals subscribe to more than one type of social network. We exploit this second-best solution in this paper and compare with-in household variation in the benefits that a household gains from subscribing to different iddir types. This approach circumvents household-level selection into different types of social networks. However, while the fixed-effect approach circumvents households unobserved factors affecting iddir membership and iddir type choice, it does not rule-out households strategic choice of different networks for different purposes. For instance, if a household subscribes to a more homogenous iddir for making friends (or social interaction) purpose while also subscribing to another heterogeneous network for insurance purposes, we cannot attribute the gains and benefits that these different networks provide as qualities of these networks. To control for strategic choice of different network types for different purposes, we rely on observational information from the household survey, which elicits households motives (objectives) for each iddir they subscribe. Thus, we start by characterizing households motives (and purposes) for joining different types of iddir associations. It is worth noting that the main purpose of iddir associations in Ethiopia is to provide funeral support for deceased (family) members of the network. Nevertheless, practically they also serve as: (i) source of small-scale loans; (ii) a means for making social interactions (socialization); (iii) a source of protection (insurance) to members; and (iv) a means for making friends. Thus, we first characterize households demand for each iddir they subscribe as a function of different observable attributes of iddirs and households by estimating the following equation: Y hi h 1 ( composition hi) 2' Xhi hi (1) 12
where Y hi stands for households demand (motive) for each iddir they join; h stands for household-specific fixed effects, 1 measures the effect of iddir composition on household s demand (motive for joining) each iddir. X hi represents a vector of household and iddir-level characteristics that may affect households demand for each iddir they subscribe. hi captures other unobserved factors that may induce heterogeneity in our outcomes of interest. In Table 4 we provide linear probability model estimation results of equation (1). In column 1 of this table we regress households demand for loan service from iddir networks as a function of observable characteristics of households and iddirs, while column 2 of this table controls for household fixed effects. Similarly, columns 3 and 4 estimate households demand for social interaction with the latter column controlling for household fixed effects. Columns 5-6, and columns 7-8 do similar exercises for households demand for protection (insurance) and making friends, respectively. Table 4: Households Demand (Motives) for Joining Iddir Associations (besides funeral support) Explanatory variables (1) Loan service (2) Loan service (3) Social interaction (4) Social interaction (5) Protection (insurance) (6) Protection (insurance) (7) Making friends (8) Making friends Majority of iddir members 0.003 0.004 0.028 ** 0.021 0.010 0.006-0.008-0.012 are from this (same) village (0.007) (0.008) (0.013) (0.015) (0.009) (0.010) (0.009) (0.012) The husband is the member 0.017 *** 0.018 *** -0.005-0.009 0.012 ** 0.012 ** -0.006-0.007 (0.005) (0.005) (0.008) (0.008) (0.005) (0.005) (0.006) (0.006) Parents used to be members 0.003-0.005-0.003-0.002 0.012 0.016 * -0.024 *** -0.012 (0.006) (0.006) (0.013) (0.014) (0.008) (0.008) (0.008) (0.010) Log (size of iddir) 0.004 * 0.004 ** -0.005-0.006-0.007 *** -0.005 * 0.006 * 0.007 ** (0.002) (0.002) (0.005) (0.005) (0.003) (0.003) (0.003) (0.004) Household characteristics Yes No Yes No Yes No Yes No Household fixed effects No Yes No Yes No Yes No Yes Constant -0.034 0.030 *** 0.404 ** 0.424 *** 0.501 *** 0.198 *** 0.110 0.079 *** (0.071) (0.009) (0.171) (0.016) (0.134) (0.010) (0.078) (0.012) Number of observations 4458 4458 4458 4458 4458 4458 4458 4458 Notes: This table estimates households motive (demand) for joining iddir associations. Columns 1-2 provide estimates for households demand for loan; columns 3-4 are estimates for households demand for social interaction (socializing); columns 5-6 provide estimates for households demand for protection while the last two columns (7-8) are estimates for households demand for making friendship. Columns 1, 3, 5 and 7 include household characteristics, while columns 2, 4, 6 and 8 include household fixed effect. Robust and clustered (at household level) standard errors are given in parentheses. Asterisks: *, ** and *** indicate statistical significance at 10%, 5% and 1%, respectively. The results in Table 4 show that households joining potentially homogenous and heterogeneous iddir networks have similar motives (demands) for the various services that iddirs provide. As expected, there exists slight variation on the demand for socialization, for which iddir 13
networks dominated by your type are more preferred for social interaction purpose than those iddir networks outside your vicinity. This is intuitive and expected, but this difference disappears when we control for household fixed effects (see column 4). However, the results in Table 4 do not preclude strategic network choice of network type. For instance, the results show that iddir networks where the husband of the household is a member are more demanded for loan and protection purposes while potentially larger iddirs (in terms of size) are associated with higher demand for loan and making friends. 11 Overall, the empirical results in Table 4 suggest that although the demands (motives) for joining different iddir networks do not significantly differ across networks with different compositional structure, it is crucial to account for (and consider) strategic network choice in quantifying the gains from social networks. Thus, to characterize the different qualities (or attributes) of iddir networks and quantify the costs/benefits of these networks we extend equation (1) by controlling for strategic network choice for different purpose (motive for joining) in such a way: Y hi h '( motive hi) 1 ( composition hi) 2' Xhi hi (2) where Y hi now stands for the costs/benefits and other measurable services are associated with each iddir a household subscribes. Other notations are as in equation (1), except the fact that motive hi now stands for a vector of household motives (demands) for subscribing to each iddir. 1 is the parameter of interest which now measures the effect of iddir composition on various outcomes of interest, including the costs and benefits to members. Household s potential costs and benefits from various iddirs can be correlated due to some unobserved household characteristics. Thus, we cluster standard errors at the household level in all estimations. 12 4. Main Results and Discussion In this section, we characterize the costs, benefits and other services that iddir members have to pay and gain from their iddir networks. Given that 99% iddir associations require regular (monthly) contributions we characterize these contributions with regard to the type of iddir associations that 11 The results also provide some intuitive associations between households behavior and motive for iddir subscription. We can see that households with better social capital are generally less likely to demand much from iddir networks beyond the very common funeral support. On the other hand, liquidity constrained households have higher demand for loan and protection, while concurrently lower demand socialization and making friends. These results are suppressed for space considerations, but they are available from the authors up on request. 12 Clustering at iddir level is also justifiable, but given that we suspect some spelling errors in iddir names, we prefer to report estimates with household level clustering. We also had standard errors clustered at village-level, but this did not affect any of our inferences. 14
households join. As discussed in Section 3, iddir members also pay some contributions at the time of funeral. Thus, we also characterize these various forms of costs associated with the alternative composition of iddir networks. We expect that more compact and homogenously formed networks are easier to run, and may perhaps require lesser contribution. We also expect that the modality of contributions may vary between those iddir networks formed by more homogenous and heterogeneous members because the former are easy to monitor through peers and hence may allow payments in-kind, a cheaply available resource that many households are able to deliver. Table 5 provides these estimates for three types of costs and contributions that iddir members pay to the pool of their iddir association. The first column models households monthly contributions as a function of iddir type as well as other detail features of iddirs and households. In column 2 we control for household fixed effects. In columns 3 and 4, we run similar regressions for cash contributions at the time of funeral, without and with household fixed effects, respectively. Similarly, columns 5 and 6 of this table provide estimates for in-kind contributions at the time of funeral, without and with household fixed effects, respectively. Overall, the results in Table 5 are intuitive and anticipated. The results in Table 5 confirm that more homogenous iddir networks are more suitable to mobilize in-kind resource, and hence can rely on in-kind contribution from members while more heterogeneous networks are more likely to suffer from this challenge, and hence may have to rely on cash contributions from members. Columns 1 and 2 of Table 5 show that heterogeneous iddir networks charge almost 1 Birr extra monthly contribution compared to those iddir networks dominated by members from the same village. This amounts to around 29% of the average monthly contribution shown in Table 2. But columns 5 and 6 of this table show a reverse effect when it comes to in-kind contribution; showing that homogenous networks may rely on in-kind contributions from members. Columns 5-6 of Table 5 show that iddir members dominated by members of the same village charge 2-2.6 Birr worth of extra in-kind contribution at the time of funeral. Interestingly, in terms of cash contributions at the time of funeral both types of iddir networks charge similar costs (see columns 3-4), which is not surprising provided that many iddirs require members to attend a funeral of their member, a process that simplifies mobilizing these cash contributions at the time of funeral. Overall, these results suggest that more heterogeneous social networks may involve potential monitoring and mobilizing costs while social networks formed by homogenous members are easier to monitor and mobilize. 15
Table 5: Monetary and Non-monetary Costs Associated with Iddir Membership Explanatory variables (1) Monthly contributions (2) Monthly contributions (3) Funeral cost (cash) (4) Funeral cost (cash) (5) Funeral cost (in-kind) (6) Funeral cost (in-kind) Majority of iddir members -0.801 *** -1.053 *** 0.094 0.094 2.639 *** 2.007 *** are from this (same) village (0.135) (0.163) (0.169) (0.158) (0.439) (0.650) Motive for joining: loan service -0.694 ** 0.240 0.306-0.255-1.498 ** -2.917 ** (0.292) (0.527) (0.638) (0.649) (0.745) -1.302 Motive for joining: socializing -0.627 *** 0.046-0.087-0.466 1.686 *** 0.157 (0.139) (0.216) (0.274) (0.574) (0.415) (0.717) Motive for joining: protection -0.925 *** 0.159-0.075-0.016 3.741 *** -0.977 (0.179) (0.320) (0.256) (0.487) (0.649) -1.289 Log (years since member) 0.209 * 0.212 * -0.479 ** -0.427 0.686 * 1.067 ** (0.111) (0.124) (0.240) (0.276) (0.410) (0.518) The husband is the member 0.649 *** 0.598 *** 0.652 *** 0.726 *** -2.381 *** -2.570 *** (0.082) (0.088) (0.134) (0.149) (0.371) (0.394) Parents used to be members 0.101 0.168-0.028-0.064 0.794 * 0.063 (0.104) (0.116) (0.174) (0.214) (0.409) (0.531) Log (size of iddir) -0.198 *** -0.159 *** -0.039 0.040-0.140-0.347 (0.049) (0.059) (0.054) (0.064) (0.187) (0.222) Household characteristics Yes No Yes No Yes No Household fixed effects No Yes No Yes No Yes Constant 4.883 *** 3.311 *** 2.989 ** 2.901 *** 1.545 3.640 ** (1.076) (0.411) (1.392) (1.095) (3.578) (1.709) Number of observations 4440 4440 4308 4308 4252 4252 Notes: This table estimates households costs associated with joining iddir associations. Columns 1-2 provide estimates for households monthly contribution as a function of iddir type and household characteristics; columns 3-4 are estimates for households cash contribution at the time of funeral while columns 5-6 provide estimates for households in-kind contribution at the time of funeral. Columns 1, 3 and 5 include household characteristics, while columns 2, 4 and 6 include household fixed effects. Robust, and clustered (at household level) standard errors are given in parentheses. Asterisks: *, ** and *** indicate statistical significance at 10%, 5% and 1%, respectively. We now turn our investigation to the benefits and services that iddir members enjoy from their iddir networks. We focus on three typically common benefits and services that iddir networks provide. The first and most common service of iddir networks is a payout (in cash) payment for members of the network when a family member dies. A second service which is given by around 44% of the iddir networks in our sample (see Table 2) is provision of small-scale loans. Third, and also common service is funeral support at the time of funeral, which commonly include food and labor service (in addition to the cash payments given as payouts). Table 6 provides estimates associated with these three types of benefits and services that iddir members enjoy from their networks. Column 1 of Table 6 runs a linear probability model for an indicator variable showing whether the household has received payouts from the iddir network in the last three years as a 16
function of iddir type, monthly contributions (household investments) and other attributes of iddirs and households. Column 2 of this table extends this empirical specification by controlling for households fixed effects. Columns 3 and 4 make similar estimations for households access to loan service from the iddir (whether the iddir provides loan service), while columns 5 and 6 estimate linear probability models for an indicator variable showing whether the household has been denied a service that the household deserves. As a follow-up to this question, we also investigate households commitment to each iddir network they subscribe by assessing whether iddir members continue as members of their iddir network in the latter three follow-up surveys. We generate this indicator variable from the latter three surveys of our data. 13 Thus, in columns 7 and 8 we also run linear probability models for households propensity of maintaining their iddir networks, with and without household fixed effects, respectively. The results in Table 6 show that in terms of payouts, households subscribing to more homogenous and heterogeneous groups have statistically comparable probability of receiving payouts. In terms of loan service, households subscribing to denser and homogenous groups have higher access to loan service than those households subscribing to more heterogeneous iddir networks. This may suggest that more homogenous social networks are easy to monitor, and hence provide loan service which can be effectively monitored through peer influence and smooth flow of information among members. This result is consistent with earlier findings in the credit literature (Huppi and Feder, 1990; Wydick, 1999; Karlan, 2007) who document that homogenous group formation facilitates peer monitoring of performance of loan repayment and saving. As expected, the results in Table 6 show that those households who stayed as iddir members for longer time and those households who joined iddir networks for getting loan service have higher access to loan service from their iddir networks. Columns 5-8 of Table 6 show that on average households joining more homogenous and heterogeneous iddir networks have comparable probability of deprival of a service they deserve and comparable probability of maintaining their iddir networks. Similarly, the amounts of monthly contributions (households investments) as well as other attributes of iddir networks have limited effect on households access to some iddir services and households decisions to continue members of their iddir network. The fact many attributes of iddir networks have limited effect on households probability of quitting iddir membership hints that these networks are potentially stable, an evidence that suggests households have limited incentive to deviate from the 13 In these latter surveys households are asked whether they continued as members of each iddir they subscribed and mentioned in the baseline survey. 17