Macrostructure and microstructure: Evidence from ... - Semantic Scholar

Macrostructure and microstructure: Evidence from overlapping village networks in The Gambia Jean-Louis Arcand∗ Dany Jaimovich† Graduate Institute, Geneva PRELIMINARY VERSION. First Draft: November 2009. This draft: February 2010

Abstract Using data recently collected in 60 Gambian villages, we study the structure and interaction of different social networks and the determinants of household centrality in each of the following networks: (i) land exchange, (ii) labor exchange, (iii) tool and fertilizer exchange, (iv) credit exchange, (v) matrimonial relationships and (vi) kinship relationships. Using a variety of measures gleaned from the Social Network Analysis (SNA) literature, we find that village-level income inequality seems to play a role in determining a network’s macrostructure. On the other hand, analysis at the micro (household) level reveals that traditional roles in the village are the main determinants of household centrality and link formation, with the latter being dominated by the effect of homophily. We also present preliminary evidence concerning the complex interactions amongst various networks, including the correlation in household centrality and the overlap of the links created in different networks. Keywords: West Africa, Social Networks ∗

Economics Unit, Graduate Institute of International and Development Studies, Pavillion Rigot, Avenue de la Paix 11A , 1202 Gen`eve. E-mail: [email protected] † Graduate Institute of International and Development Studies, Economics Unit. Avenue de la Paix 11A , 1202 Gen`eve, Switzerland. e-mail: [email protected]

1

“In The Gambia, virtually everything is lendable and at times will be lent. This includes nearly all factors of agricultural production: land, labor, livestock, seed, fertilizer, pesticides, and farm tools. Craft tools, vehicles, and household goods are also lent” (Shipton (1990)).

1

Introduction

In his pathbreaking article about the embeddedness of economic interactions, Granovetter (1985) pointed to a great divide in the economics literature of the time. On the one hand, the undersocialized view of neoclassical models, in which agents are treated as atomized entities where social structure plays no role. On the other, the oversocialized view of new institutional economics, in which institutionalized social structures dominate human actions. The real world must then lie somewhere between the two extremes, with rapidly changing social structures defined by apparently unstable interactions and multiple links among agents, thereby creating dynamic and overlapping mesostructures. During the past two decades, the economics literature has advanced in terms of filling this gap, particularly with the development of game theory, behavioral economics, information economics and, especially, through the study of social and economic networks. Social Network Analysis (SNA) has been an important component of sociological research for more than a century, but economists have only recently started to incorporate it into their toolkit1 . While the concept of social capital proposed by Putnam (1995) called attention to SNA, it was the modeling efforts of Jackson and Wolinsky (1996) and Bala and Goyal (2000) that created the bases for the incorporation of networks in economic analysis. Since then, a growing body of theoretical literature has been developed, accompanied by an increasing number of empirical studies that address issues as diverse as job search, criminal activities and the use of the Internet2 . The present study aims to contribute to the recent empirical literature on the study of social networks in rural economies. This is the first step in a research agenda related to very detailed data about social and economic interactions recently collected in 60 Gambian villages, in the context 1

In Granovetter (2005) there is a broadranging summary of the sociological point of view concerning the influence of social networks on economic activity. 2 For a complete review of the history and recent progresses of SNA in economics, see Jackson (2009).

2

of the impact evaluation of a randomized Community-Driven Development Program (CDDP). The empirical analysis that we present here is meant to provide a description of the networks and should not be interpreted in causal terms. The paper is organized as follows: Section 2 briefly summarize the relevant literature on social networks in the field of Development Economics. Section 3 explains in detail the process of data collection and presents descriptive statistics on villages and households. In section 4 we describe the networks, the measurements for their characteristics and a first set of results for determinants at different levels. Section 5 studies the manners in which overlapping networks interact, while the final section concludes.

2

Literature review: empirical analysis of networks in Development Economics

In Development Economics, empirical studies of networks have been mainly confined to various forms of sharing arrangements and the diffusion of information. Concerning the former, there is a massive corpus of theoretical work on sharecropping from the 1970s and 80s (a summary of the theories in Singh (2000)), and important recent empirical studies such as Fafchamps and Gubert (2007) and De Weerdt and Dercon (2006) that have shown the importance of risk-sharing arrangements, while labour-sharing arrangements have been singled out by Krishnan and Sciubba (2009). In terms of social networks as the key factor in the transmission of information for technological innovations in agriculture, significant contributions have been provided by Conley and Udry (2008) and Bandiera and Rasul (2006), while Miguel and Kremer (2003) consider the use of various drugs in network-theoretic terms. However, few studies have tried to understand the joint determinants and interaction of the different networks that overlap in rural economies. To our knowledge, the first empirical attempt to deal with the issue is Udry and Conley (2004). They use data stemming from 4 villages in Ghana to relate four different networks: information, credit, land and labour, finding some preliminary evidence of network interconnection. Udry and Conley (2004) also explore the determinants of link formation in the different networks, an exercise that a growing number of studies have

3

implemented recently thanks to the development of dyadic regressions techniques. Examples include Arcand and Fafchamps (2008) who consider the determinants of participation in Community Based Organizations in several hundred villages in Burkina Faso and Senegal, and Comola (2008), who examines networks of contacts for potential help in the case of an emergency in a village in rural Tanzania. Comola (2008) is also related to our study in terms of expanding the concept of a network from a mere collection of dyadic interactions to a more careful analysis of network “architecture”, emphasizing that the position of individuals in the network plays a fundamental role in link formation. Similarly, studying labour-sharing arrangements in Ethiopia, Krishnan and Sciubba (2009) note the importance of differences in the symmetry of networks in terms of economic performance, and show how differences in endowments of participants are key determinants of the structure of the network.

3 3.1

Data and setting Data collection strategy

The methodology adopted for the present study differs from that of traditional household surveys in which a random sample of households is collected in each village. Our goal is to collect data for the full range of nodes participating in different networks within the village and the external links of these nodes. Given the costs associated with implementing LSMS-type surveys for all households in a village, we instead used structured group interviews geared to collect quantitative information. We therefore carried out village censuses through gatherings co-organized with the village chief (Alkalo) and district-level officers. In such village meetings it is possible to obtain relatively coarse quantitative information -with a particular focus on socioeconomic interactions- for almost all households in each village. This type of approach is common in ethnographic research and is related to the rapid rural appraisal methodology that has been successfully used in the past for quantitative analysis in different disciplines (Chambers (1994)). The data collection consisted of the following strategy: Two enumerators spent several days in each village, 1 to 2 days per village were usually sufficient. The first step was a meeting with the village chief or his representative, to explain the propose of the visit and to offer a traditional symbolic gift to the village (kola nuts). The Alkalo was consulted about general village 4

information and we secured authorization to use the taxation roster, where all compound heads are registered, to create a list of households. With the Alkalo’s assistances, household heads were invited to gather in the village’s meeting point (bantaba). As the heads were arriving to the bantaba, enumerators formed groups of 5 to 8 people and administered simultaneously the questionnaire to all members of this group. Though questions were closed-ended, we allowed for some interaction among participants in order to clarify data that were incomplete or doubtful. The process was repeated until all households had been interviewed. Sometimes it was necessary to move to some specific compounds to individually administer the questionnaire because of households head being sick and (in a few cases) because some individuals refused to be interviewed in the bantaba. Simultaneously in some villages Focus Groups were conducted for three groups of villagers, men, women and youth, with the goal being to better understand the structure of each community. The structured group survey consisted of two categories of information. The first section was a standard (and very lean) household questionnaire designed to collect a vector of household characteristics including: economic and demographic indicators, household head characteristics, extent of ties with traditional village authorities, membership in various village-level groups, household migrants, and sources of external information. In order to see how such data differs from that gathered in standard surveys, Figure 1 compares the distribution of income in 15 villages where, some months earlier, a random sample of households were interviewed using LSMS techniques in the context of the baseline analysis of the CDDP evaluation, with very detailed questions about different sources of income for the month previous to the survey. Both distributions share the same support, but group data is more concentrated at the mean and has a notably high kurtosis. It is possible to see that in the left-hand portion of the densities, where poorer households are located, the distributions are very similar, but richer respondents seems to provide downward biased information in a group survey. The second section of our survey instrument was specifically designed to understand the social networks in the village. Each household head was extensively asked about the interaction with other inhabitants of the village and about the existence of external connexions in six dimensions using the following questions: • LAN D=Of the land you cultivated last year: did you lend out or borrow in land from other villagers? If yes, what is the amount of land? 5

• LABOU R=Did you, or any members of your household, worked for other households during the last year (2008-9)? If yes, how many days? • IN P U T =Did you lend out to or borrow in any means of production (such as tools or fertilizer) from other household in last year (2008-9)? • CREDIT =Did you lend out to or borrow in money from other household in last year (2008-9)? • M ARRIAGE=Have any of your households members married members of other households? ({+} if your family increased in size as a result of this, {-} if it decreased) • KIN SHIP =With which households do your family members have kinship relationships? ({1} if the household head has a direct blood relationship (father/child, direct brother/sister), {2} if the wives of the head have a direct blood relationship) with this information we built “matrices of interaction” reflecting the social networks in the village. Given that we systematically established a list of all households in the village before the gatherings, we were able to assign a unique identifier to each household head beforehand. We then asked for the names of all the villagers with whom the household head had had interactions. In Figure 2 we present an example of a network, the land market “matrix” for the village of Passamassi Fula. The data presented in Figure 2 give the number of hectares transacted, with a minus sign indicating lending and no sign indicating borrowing.

3.2

Sample and household data description

The survey was conducted between February and May of 2009. 60 Gambian villages, mainly in rural areas (just 4 villages are in semi-urban areas), were randomly selected. In order to achieve a modicum of representativity at the national level, for each of the 6 Local Government Areas (LGAs) eligibles for the CDDP project, out of the 8 in which Gambian territory is divided, we randomly selected wards, a smaller geographical division that tends to be homogeneous in geographical but heterogeneous in socio-cultural terms. Finally, 3 to 8 villages were randomly drawn within each ward. Our sample consists of 3,320 households. We were particularly careful to minimize selection problems by attempting to interview all households in the 6

village, yielding a median household coverage rate of 94% (when the number of surveyed households is compared with those on the taxation list). If a household head was absent or not able to answer, we allowed someone else to answer for him (or her), usually either another household member or a close relative. This was the case for 28% of the sample. In Tables 1 and 2 we present descriptive statistics for the data collected using the first survey instrument, the general questionnaire. In the first of these tables we present village-level statistics (60 observations) while in the second we present the household-level data (3,320 observations). The data on village population available in the 2003 Gambian Census (and which we used to construct our sampling frame) turned out to be significantly out of date and, in some cases, very inaccurate. This explains that while we were targeting villages with populations between 300 and 1,000 inhabitants, in practice the smallest village has 202 inhabitants and the largest 1,402. The average population of the surveyed villages is 586 inhabitants. Population density, at least when the denominator is the inhabited area, is high, with an average of 6,900 inhabitants per square kilometer. In contrast, agricultural land was usually very abundant. The average amount of agricultural land per active worker was around two hectares, when land usage rights for the year of the survey were considered, with a great deal of variation given that the average standard deviation at the village level was 4.6 hectares. Average household size, on the basis of the household data, is 12.6 members (see Table 2), and slightly lower (11.6) when we considers the mean of village medians reported in 1. While some households appear exceedingly large, we found that respondents were very clear in terms of their definition of a household. The presence of households with more than 50 members (approximately 1% of the sample) is explained both by the polygamous nature of Gambian rural society and the existence of marabout households where the household is constituted by a mass of disciples and other followers. 45% of households declare to be polygamous and there was at least one marabout in half of the villages in our sample. As is to be expected in West Africa, a very small number of household heads are females (7%), generally represented by widows and mainly concentrated in the semi-urban areas on the outskirts of Banjul (the national capital). These villages also accounted for the few non-Muslims in the sample (4%). Our sample is representative of the ethnic diversity in The Gambia. 41% of the respondents are Mandinka, 26% Fula, 8% Wollof, 7% Jola and 4% 7

Serer, with the rest either belonging to local ethnic minorities or being nonGambian (these represent 4% of the sample)3 . While some villages are very homogeneous in terms of ethnic composition, others are very diverse. We P 2 build an Ethnic Diversity Index as EDI = (1 − N e=1 se ), where se is the share of each ethnic group in the village. The EDI ranges from zero (complete homogeneity) to 0.84, with a mean of 0.3. The economic conditions in the villages in our sample correspond, by and large, to those of traditional rural societies. There is almost no access to electricity, with an average village-level access rate of 3%. 88% of households have no access to an improved source of water, while 38% lack access to a private toilet. 38% of household dwellings are built with grass. 82% of the respondents declared having no formal education, although a substantial fraction of the villagers received some kind of koranic education and usually master basic Arabic language skills. The main economic activity is usually related to agriculture (66% of households have this as their main activity) or fisheries (6%). Nevertheless, a Herfindahl index of sectoral heterogeneity shows a significant degree of diversity, driven mainly by the presence of inhabitants working outside the village (25%). Monetary income is very low. The average (self-declared) annual income per capita is 3,565 Gambian Dalasis, which corresponds to approximately 90 Euros (using the exchange rate of 2009), and just around 12% of this income stems from agricultural activities. The distribution of income is, however, not necessarily egalitarian in all villages. Though the average Gini coefficient is 0.34, it reaches 0.5 in some villages. The higher level of inequality reported in some villages seems to be driven by remittances and off-farm jobs. Around half of the respondents declare to have current or previous household members who work outside the village, including 12% who receive remittances from overseas migrants outside Africa.

4

Social and economic networks

4.1

Cultural anthropology of the networks

We now turn to the description of the main data we collected for the present study, that on village economic and social networks. Our basic unit of study will be the household, the definition of which is not straightforward. Villagers 3

Compared with the data for rural areas of the 2003 Census for the Gambia, our sample overrepresents Mandinkas (33%) and underrepresents Wollof (15%).

8

in rural Gambia are usually organized into compounds, a concept which corresponds to members of the same family, usually related by blood or marriage, living in a group of huts surrounded by a grass fence. Usually a compound can be identified as a household, meaning that the members eat together, organize daily activities in common and have a head who takes most important decisions. Nevertheless, it is sometimes the case that some members of a compound declare themselves to be an independent household, with their own head. The independence of a household within a compound is related to two local concepts: the dabadas or farm production units and the sinkiros, cooking and consumption units (Webb (1989)). People in the same dabada tend to belong to the same sinkiro, and more than one of these units can be found inside compounds, particularly the largest. The intra-compound distinction was, in most cases, clear to the alkalo and all other village inhabitants. Just 16% of the households heads in our sample are not heads of the compounds where they live. The choice of a household as the unit of analysis is not without cost. As noted by Udry (1996) and Goldstein and Udry (2008), among others, the assumption of within household Pareto efficiency is far from reality in West African rural societies. In the context of The Gambia, Carney and Watts (1990) and von Braun and Webb (1989) have shown that the decision-making process inside compounds and households is complex, particularly considering that individual resources allocation is sometimes divided between common farm land, known in the Mandinka language as maruo, and private plots or kamanyango, where members of the household, such as some of the wives or the young people, can manage a crop under his/her control. We asked how much of the respondent’s land is a kamanyango (or any similar concept of private plot in the local dialect). Just 6% of the compound heads that have cultivated a plot declared a significant portion (more than 25% of the total) of kamanyango and 19% around one quarter of the land 4 . The relationships that we are uncovering with our data collection strategy might therefore be interpreted as the networks centered on the household head and might not necessarily reflect the complete network of the household. Carney and Watts (1990) describe the clear division of labour in Mandinka society, which is largely similar in villages dominated by other ethnic groups. Men cultivate cash-crops plots, mainly groundnuts, but recently also fruit trees, in conjunction with some staple food such as millet, maize and sorghum. 4

Of those declaring no kamanyango, a portion was not familiar with the concept and was not possible to explain its meaning.

9

Women grow rice and take care of the village garden (if there is one)5 . In semi-urban areas there are almost no agricultural plots and inhabitants usually have titles for the land they use. This is in contrast with rural areas, where there is usually no land scarcity (at least in terms of quantity, but not necessarily in terms of quality) but the unwritten rights over its usage are determined by the descendants of the village’s founders, generally the Alkalo and his direct relatives. In some cases, the kabilo (clan) heads that might not be related to the founder’s lineage but represent diverse kins of settlers are entitled to permanent usage rights6 . All other villagers must borrow plots on either a seasonal or an annual basis from them, in agreements that can also last for several years (Chavas, Petrie, and Roth (2005)). Sometimes other individuals own small plots of land outright that can be lent or rented. The transactions in the LAN D network will therefore principally be given by a moneyless assignment of plots, though they can sometimes imply a remunerated rented contract. While land is not scarce, labour is, particularly before and during the rainy season. The solutions to a household is shortage of workers constitute the core of the interactions in the LABOU R network. The main options are (Swindell (1978)): the use of kafos, an organized workforce of villagers who participate in the provision of public goods but who can also be hired for a fixed wage by a household; the hiring of particular villager or outsider at a given “market” wage; inviting other villagers or outsiders to help with household tasks in exchange for various kinds of goods, labour or even a marriage arrangement; the use of strange farmers that provide part-time labour in exchange for the right of use of part of the family plot for his own benefit. The links in the IN P U T network are defined as exchanges of means of production that imply a monetary or opportunity cost for the lender, such as tools, cattle, fertilizer, seeds and the like. While cattle are usually lent for milk, manure and transport during the whole year, agricultural tools and inputs are particularly important in the period before the rainy season. 5

This division of labour is very ingrained. Several efforts of the British colonial authorities and post-independence programs have failed to incorporate men in rice production, a very important factor when it comes to understand the country’s dependence on imported food. 6 As remarked by Webb (1989), the rights over land are related to the old social structure, with the former highest castes having the most productive plots. Apart from its historical consequences, mainly with the division of freeborn and former slaves, the caste system is not relevant in rural Gambia nowadays, and interviewees were generally not willing to talk about their ancestors’ caste.

10

This seasonality is also explained by the investment decisions of the villagers. While some invest in means of productions after the crop season, others do not, either because of a bad year for cash crops or due to substitution effects driven by consumption of non-productive goods. The exchange can also be driven by risk-sharing behavior with households investing in differentiated means of production. Information concerning the CREDIT network must be taken with caution. While there exists lively money lending and borrowing amongst villagers, it is not easy to secure these data. During the field tests prior to the survey we were initially reluctant to ask information about money exchange, thinking that might be perceived as being disrespectful. On the contrary, we found that villagers were in general willing to respond to questions related to credit. The clue was giving by one of our local enumerators: “In Islam there is no interest rate. If you lend money it means that you are helping at the moment when the other really needs it, so you are doubly blessed. While usually lenders will not reveal the information, grateful borrowers will.” On the other hand, this idea can be in contradiction with the local concept of juloo (rope): “It can refer to a small-scale trader, or to credit or debt. Every Mandinko knows the meanings are related. Traders are also lenders, and their loans, while sometimes useful like a rope ladder, also tie down a farmer like a rope around the neck ” (Shipton (1990)). Apart of the direct borrowing of money from another household in the village, there is also the possibility to obtain credit from external sources, both informal and formal (mainly rural development banks or microcredit agencies), or from some village-level rotating saving and credit associations (ROSCAs), locally know as osusus 7 . Other forms of organized saving, such as the village bank where women save incomes from their gardens or skill center are also available. In the small villages in which we conducted our surveys kinship relationships are very common and are usually dominated by the lineage of the village founders and the oldest settlers. Intermarriage within families is common and there is also a very active inter-village exchange of brides. In polygamous rural Gambia, is often the case that the first marriage is arranged between the elders of the respective families, while the other wives are requested by the husband according to his capacity to sustain a bigger family. 7 In the CDDP baseline survey, 58% of the rural villages declared to have at least one osusu association.

11

4.2

Network measurement

We will consider a household as a node i in each of the m networks g m . We restrict each network g m to consist of a set of nodes belonging to N = 1, ..., n where n is the number of households inside each village. Figure 3 present a graphical representation of a network (built using the data shown in Figure 2), where the arrows represent the existence of a directed link between nodes and the nodes in the upper left section do not have any links in the network. A basic representation of the embeddedness of a node i in a networks g m is its degree, the cardinality of its neighborhood as measured by the number of links with the other nodes j: di (g m ) =

P

j `ij ,

where `ij = 1 if there is a link between i and j and zero otherwise. In the case of our directed networks this is an out-degree measure. In the first column of Table 3 we present basic statistics for the distribution of degree in (g m ) . different networks, expressed as a fraction of the total possible links din−1 The average degree tends to be around 2%, except for kinship with 9%, but the data are very heterogeneous, indicating important differences in the centrality of households. Though we defined the networks in terms of within village interactions, we also registered the links of each household with nodes outside the village. Columns 2 and 3 of Table 3 present the distribution of the percentage of nodes with at least one external link, where external-in are links created to bring something to the village, with external-out representing the opposite direction. External actors are very important for the marriage network, while for economic networks they represent less than 10% of the total (exept for CREDIT ), lending support to our idea that the most important interactions in the networks take place inside villages. The second panel of Table 3 presents village-level, or macrostructure, descriptive statistics for the networks. The number of links corresponds to P m i di (g ), from which we can build a measure of density of the network, defined as the average number of links divided by n − 1. Densities are around 1% for land, labour and credit, 2% for inputs and marriage and 7% for kinship. An important regularity observed in the SNA literature is the very high 12

density of local neighborhoods, related to a lightly clustered link formation (“my friends tend to be friends amongst themselves”). A clustering coefficient Cli (g m ) can be used to study this feature, calculating the local density for each node i as: m

Cli (g ) =

P ` ` ` j6=i;k6=j;k6=i ij ik jk P , ` ` j6=i;k6=j;k6=i ij ik

and the clustering coefficient for a given network g m can be calculated as a weighted average of Cli (g m ): Cl(g m ) =

X

Cli (g m )wim ,

i

wim

is a weight related to the size of node’s neighborhood. The level where of clusterness varies significantly among networks. In the case of land, no significant cluster effects are apparent, while (almost by tautology) the kinship network is highly clustered. The other networks present non-negligible levels of clusterness. The analysis of a network will differ between situations in which only dyadic or tryadic interactions exist (in which case dyadic regression analysis will be sufficient) and situations in which more complex structures connect a group of nodes. The latter situation corresponds to the case of a significant number of nodes connected amongst themselves, meaning that there exists at least one path connecting the nodes that are part of this structure. This is the basis to define a component cm as a subnetwork of network g m , such that the nodes of cm are connected amongst themselves but disconnected from the rest of the nodes in g m . We create an index of compactness (this is not standard in the SNA literature) defined as: m

Cmp(g ) = 1 −

P m P c m i

di (g )

,

where cm is the number of components in g m . Thus, if the network is mainly dyadic in nature, the index will be close to zero, while the closer the network is to having just one component, the index will approach 1. On average, it is clear that networks in rural Gambia are very compact, though they are subject to considerable heterogeneity, reflecting that some villages display sparser relationships than others. P

13

4.3

Determinants of network architecture

We now turn to the analysis of the village characteristics associated with network architecture. Our approach will be to take each network gvm as the unit of analysis, with m = {LAN D, LABOU R, IN P U T S, CREDIT , KIN SHIP , M ARRIAGE} and v corresponds to each of the 60 villages for which we have information, yielding a total of 360 observations. The empirical specification will be as follows: ymv = αward + αm + αethnic + αactivity + Xv βv + mv

(1)

where ymv is one of the three village-level network characteristics de2 scribed in section 4: density, clusterness and compactness, mv ∼ N (0, σmv ) is the error term, clustered at the village level, and Xv the vector of relevant village characteristics. We include a set of fixed effects: taking advantage of the stratified nature of our sampling scheme, we include ward-specific effects (αward ) that control for several geographical characteristics, such as distance to the capital and other important population centers or climate. We also include network-specific effects (αm ), as well as dummies corresponding to the predominant ethnic group in the village (αethnic ) and its main economic activity (αactivity ). In table 4 we present the results when we estimate over the pooled set of networks, as well as when we divide our sample into economic networks (LAN D, LABOU R, IN P U T S and CREDIT ) and family networks (KIN SHIP and M ARRIAGE). Several different specifications were estimated (available upon request) though we confine our reporting to specifications that were robust to changes in the included explanatory variables. A first (and probably obvious) result is that network density is decreasing in village population, while this village characteristic does not appear to be a significant determinant of clusterness and compactness. When the average household size in the village is considered there is a significantly positive effect on network density (again an unsurprising result, since larger households will have more members who can potentially interact with other households in the village). Larger households also yield significantly more clustered networks, particularly when the network in question is economic in nature. Income inequality, as measured by the Gini coefficient constructed from self-declared household monetary income, seems to be an important factor in increasing both density and compactness. This is an interesting result that may be driven by the fact that differences in endowments provide in14

centives to create more links but less components, an idea related to the hypothesis of Krishnan and Sciubba (2009). Other measures of heterogeneity inside the village also seems to be relevant. The ethnic diversity index is negatively related to density, providing some evidence of a reduction in link formation when different ethnicities coexist in the same village. There is no effect of these diversity variables on clustering, but an index of the diversity of household economic activity associated with a significantly lower level of clusterness. Among the various indicators of poverty, the lack of access to electricity and a private toilet (the indicator which displays the most variation amongst villages) increase density, and the former also compactness. When monetary income per capita is considered, the only effect is a reduction in clustering for poorer villages. Another result that appears to be robust (at least in sign, but not always in statistical significance) is that the percentage of female household heads negatively influences density and clusterness. Semi-urban village networks are denser, a result that seems to be driven the fact that marriages within the village are more common there. A final result concerns the importance of interactions outside the network, measured as the percentage of households with external links in each villagenetwork. Though such relationships are seldom statistically significant, they do appear to exert a negative effect on compactness.

4.4

Determinants of household centrality

In this section we explore the micro characteristics of networks, using households as our unit of analysis. The main goal is to study the determinants of each ego i’s centrality in each network gvm , as measured by its degree di (gvm ). We estimate a system of seemingly unrelated regressions (SUR) for the m = 6 networks, allowing for the interdependence of the networks in terms of statistical inference. The model to estimate is as follows: m m m m + Xrole βtrad + m di (gvm ) = αvm + αethnic + αacty + Xhh βhh iv

(2)

where we include fixed effects at the village level (αv ), as well as for the household head’s ethnic group (αethnic ) and economic activity (αacty ). Three groups of explanatory variables are used: Xhh is a vector of general household characteristics, Xrole is a vector of characteristics related to the role of 15

the household head in the village. In Table 5 we present the results for the Xhh vector of regressors. The number of observations has decreased to 2784, from the original 3320, due to lack of information on some variables at the household level. Again, household size plays an important role in how active a node is in networks, and has a positive effect on degree for most of the networks. While people declaring having some level of formal education are not more central, those with Koranic studies are. Age, gender and religion of the household head, and other household characteristics, seem not be to important determinants when controlling for other variables. For the economic networks, we find evidence that decisions are taken at the household level, with the exception of land exchange, because compound heads are not significantly different in degree compared with others. This is not the case for family networks, where compound heads are indeed more central. In terms of monetary income, we divide households into quartiles: relative income appears to have no effect on a household’s degree of centrality, with the exception of the middle quartiles in the input network. This result remains unchanged when we replace the quartile dummies with income per capita per se. In Table 6 we show that a crucial determinant of household centrality is the role play by the household head in the village. The village chief or Alkalo always has a highest degree, with the exception of the marriage market. In rural areas, the Alkalo is the oldest descendants of the village founders and usually the most respected figure as well as the one that takes the main decisions at the village level, usually in consultation with the Council of Elders. While the Alkalo’s relatives do not necessarily have a higher degree (not surprisingly since on average 30% of the households are blood-related with the chief, sometimes increasing up to 90%), his assistants do. For the Council of Elders, only the head is more central, both in the family networks and in terms of labour exchange. Another important institution is the Village Development Council (VDC), the most decentalized part of the Gambian national system of development coordination, that links a village with officers at the ward and LGA levels. The members of the VDC are representatives from each kabilo, each Community-based organization (CBO) and other co-opted members. All VDC members are very central in terms of kinship and tend to be central in the LABOU R and IN P U T S networks, but the head of VDC has the highest degree and is also important in the CREDIT network. Nevertheless, only the Alkalo is the center of the LAN D network, reflecting his traditional 16

possession over the village’s territory. In half of the villages we found a villager trained to provide first aid and birth assistance, the village doctor, who is very central in the family networks but not in the economic ones, with the exception of IN P U T S. We also asked if respondents had treated other villagers using traditional methods, since the traditional healer is a very respected person, a fact reflected by his highest degree in some of the networks. The condition of religious leader seems not to be preponderant in the networks we are studying, once one controls for other variables. While all villages have at least one Imam that leads prayers, just a half have a Marabout, a respected Koranic teacher who is sometimes imbued with mystical powers and who maintains some syncretic pre-Islamic traditions such as making amulets for good luck. Imams have higher degree in LABOU R and lower degree in CREDIT while Marabouts appear not to participate. Other roles such as that of the traditional musician and storyteller (the Griot) and the officers outside the village are not significantly more central.

4.5

Determinants of link formation

Theoretical predictions of link formation are difficult. Applying the concept of pairwise stability proposed by Jackson and Wolinsky (1996) usually results in multiple equilibria and unclear predictions. In this paper we will limited ourselves to the empirical estimation of the observed links, assuming that: +j m +j m m `m ijv = 1 if B[di (g )]−C[di (g )] > 0, `ijv = 0 otherwise, m m where B[d+j i (g )] is the benefit of increasing i’s density in network g m by creating a link with j and C[d+j i (g )] is the associated cost. The benefits are (not exclusively) given by the access to the resources involved in each network (for borrowers) and to well connected people via link creation with households that play important roles in the village. The costs are associated with losses in resources (for lenders), with the opportunity cost of creating a different link, and are also a function of the differences between i and j along a number of metrics, including household size, ethnic group, kinship lineage and educational attainment. We will follow the literature on dyadic regressions by posing the following empirical specification:

m m m m `m ijv = αv + |Xiv − Xjv |βdif + (Xiv + Xjv )βsum + wijv βdyad + ijv

17

(3)

where the vector X is made up of household demographic characteristics (Xhh ) and their roles in the village (Xrole ). We study the undirected formam tion of links, which implies that `m ijv = gjiv and, to preserve symmetry on the right-hand-side, we follow Fafchamps and Gubert (2007) by specifying three types of regressors. The coefficient βdif associated with the distance in attributes between i and j. The coefficient βsum associated with the sum of the attributes of the members in the dyad. The variable wijv corresponds to characteristics that are common to each member of the dyad. Whenever an element of X is binary we include two dummies, one equal to unity if (xiv , xjv ) = (1, 1) (both share the attribute) and another equal to one if (xiv , xjv ) = (1, 0) or (xiv , xjv ) = (1, 0) (just one has the attribute). The disturbance terms m ijv are allowed to be correlated across observations involving the same individual using the two-dimensional clustering methodology described in Cameron, Gelbach, and Miller (2009). We present the results divided into the two vectors of regressors, Xhh and Xrole . The former are presented in Table 7. As expected, and in line with previous studies (e.g. Arcand and Fafchamps (2008), Fafchamps and Gubert (2007)), sharing the same ethnicity increases the probability of two households matching in various networks, though not for LAN D and CREDIT . Others similarities also matter. When both households in the dyad declare having high quality plots, and the bigger the sum of their land holdings, there is an increase in the likelihood of a link in all networks, something that might be taken as first evidence of an “elite network”, because the best land usually belong to the first settlers and kinship linages of the upper classes. The more important agriculture as a source of income (measured as the share over total income), the greater the probability of exchange in LABOU R and IN P U T S, while a link in the CREDIT market is more likely when the sources of income are more differentiated (which might be taken as preliminary evidence of some sort of risk-sharing). Households with formal education and high (self-declared) incomes are also more likely to create a link for CREDIT . While the age of the household head seems to be important only for family networks, household size matters for all networks, both when the total number of members and the number of workers are considered. For some networks, difference in size also counts, with links created between larger and smaller units. Differences in education are also important, usually reducing the probability of a match, except in the case of CREDIT . The dyadic effects of village roles are presented in Table 8. When the Alkalo is present in the dyad, the probability of a link always increases, something that happens sometimes (but not for all networks) for VDC mem18

bers, the Alakalo’s assistants and Elders. Imams also tend to create more links, with the notable exception of CREDIT . We again find evidence of the “elite network” effect, with the Alkalo’s relatives being less likely to form links with other villagers. Overall, we can say that link formation is determined by the roles that household play in the village and by similarities amongst households (homophily), while differences between households are important in a relatively small number of specific cases.

5

The interaction of networks

In this section we attempt to shed some light on the the patterns that underly the complex interactions among the different networks we are studying. Figure 4 provides a graphical representation of the complexity that can exist when overlapping networks are considered, even in a very small village, such as Pasamassi Fula, with just 30 households. When each network is considered separately, there are various isolated nodes (as in Figure 3), but the whole village is connected when all economic networks are considered at the same time, with several different paths connecting the nodes and multiple links per node, links that sometimes overlap but in several cases create network specific matches.

5.1

Centrality across networks

We first analyze whether a household’s centrality is similar across different networks. A first approach is to correlate a node’s degrees in each network, as shown in Table 9. Households that are active in some networks tend to be so in others as well. All correlations are positive and significant at the 1% level of confidence. Though suggestive, the results from these simple correlations must be interpreted with caution because of the multiple common determinants of the degree of centrality. Ideally, degrees would be estimated as a system of equations as follows: m m m + αacty + Xβxm + Xext βext + m di (gvm ) = di (gv−m )ψ m + αvm + αethnic iv

(4)

where di (gv−m ) is the vector of a household’s degrees in the different networks excluding di (gvm ), and X = [XHH , Xrole ] is the vector of household characteristics. The problem is to find exclusion restrictions to identify the 19

m system. One possibility is to use Xext , the external links of each household i m in network g . In Table 10 we explore this idea by expanding equation (2) to external links by including a dummy variable which is equal to 1 if the household has external networks, either bringing or giving something/someone to the village. We denote these variables by dextin and dextout respectively. Interestingly, an ego’s degree appears to be influenced both by the existence of external links in the network in question but also in the others. For example, in the land network, external land transactions in both directions yield a fall in an ego’s degree, but external transactions in labour, importing inputs and giving credit have the opposite effect. We take this effect as a first piece of evidence concerning the interconnection of these village networks.

The results in Table 10 provide evidence that using external links as exclusion restrictions to estimate the system of equation (4) are not a good option. As such, we will limit ourselves to estimating each equation in (4) separately. The results are presented in Table 11, where we restrict the reported coefficients to those which correspond to di (gv−m ). The positive relationships between various degrees hold, though some of the conditional correlations become statistically insignificant at the usual levels of confidence. This is the case for CREDIT with KIN SHIP , LAN D and LABOU R, M ARRIAGE with LAN D and LABOU R, and KIN SHIP with LAN D. These results are in line with the idea that a node’s centrality can change across networks.

5.2

Matches across networks

Apart from the possible overlap in terms of centrality, it is important to investigate whether actors who create links with some villagers in one network maintain the same pattern in other networks. To explore this idea, we first calculate the correlations between the vectors of partners for each household head across networks. The results are summarized in Table 12, where all mean values are positive, but generally not statistically significant. If we take the median, it is possible to see that for all interactions, except LAN D-IN P U T and M ARRIAGE-KIN SHIP , at least half of the values are negative, a result that suggests an important degree of independence among the different networks. Nevertheless, this does not necessarily mean that the probability of establishing a link between i and j will not be affected by the fact that both also have a link in another network. In order to study this idea, we expand the dyad equation given by (3) in the following way: 20

−m m m m m m `m ijv = αv +gijv +Xext(i+j) βext +|Xiv −Xjv |βdif +(Xiv +Xjv )βsum +wijv βdyad +ijv (5) −m where the vector `ijv are the potential links in the dyad formed by i and j in village v in all the networks except in m. `−m ijv = 1 if the link exists, and is equal to zero otherwise. We also include the sum of external links of both members of the dyad (the differences in external links were never significant in the estimations), including dextin and dextout . In Table 13 we first present the results for the external links, which are only important in terms of reducing the probability of a match in LAN D and IN P U T S, while there is little evidence of interaction among networks. In Table 14 we present the results for the estimation of the probability of link formation given the existence of links in the other networks and find that the existence of a link in any network increases the probability of a match in all the others, even after controlling for a series of characteristics of the members of the dyad.

6

Conclusions

In this paper, we have investigated the determinants of network structure in a sample of 60 Gambian villages using unique data that we recently collected. Analyzing network architecture at the macro (village) level, we find that income inequality was an important driving force in determining both network density and network compactness. Household-level analysis revealed that income per se played a limited role in determining an ego’s degree of centrality, with traditional social roles playing the predominant role. These social roles are also crucial in explaining the creation of links, alongside homophily. Finally, preliminary evidence points to the existence of complex interconnections amongst a given household’s various networks, with the existence of a link for a dyad in any network increasing the probability of links in the others. The results of this paper should be taken as a description of network determinants and the manner in which networks overlap, and should not be interpreted in causal terms. Further research on this particularly rich dataset will be geared towards extending the structural estimation of the network characteristics and in understanding the directionality of and causality behind link formation, thereby hopefully leading to a deeper understanding of the determinants of network formation in poor West African villages.

21

References Arcand, J.-L., and M. Fafchamps (2008): “Matching in CommunityBased Organisations,” . Bala, V., and S. Goyal (2000): “A Noncooperative Model of Network Formation,” Econometrica, 68(5), 1181–1229. Bandiera, O., and I. Rasul (2006): “Social Networks and Technology Adoption in Northern Mozambique,” Economic Journal, 116(514), 869– 902. Cameron, C., J. Gelbach, and D. Miller (2009): “Robust Inference with Multi-way Clustering,” University of California at Davis, Department of Economics. Working Paper 09-9. Carney, J., and M. Watts (1990): “Manufacturing Dissent: Work, Gender and the Politics of Meaning in a Peasant Society,” Africa: Journal of the International African Institute, 60(2), 207–241. Chambers, R. (1994): “The origins and practice of participatory rural appraisal,” World Development, 22(7), 953–969. Chavas, J.-P., R. Petrie, and M. Roth (2005): “Farm Household Production Efficiency: Evidence from the Gambia,” American Journal of Agricultural Economics, 87(1), 160–179. Comola, M. (2008): “The Network Structure of Informal Arrangements: Evidence from Rural Tanzania,” PSE working paper no 2008-74. Conley, T., and C. Udry (2008): “Learning About a New Technology: Pineapple in Ghana,” forthcoming American Economic Review. De Weerdt, J., and S. Dercon (2006): “Risk-sharing networks and insurance against illness,” Journal of Development Economics, 81(2), 337– 356. Fafchamps, M., and F. Gubert (2007): “The Formation of Risk Sharing Networks,” Journal of Development Economics, 83(2), 326–350. Goldstein, M., and C. Udry (2008): “The Profits of Power: Land Rights and Agricultural Investment in Ghana,” Journal of Political Economy, 116(6), 981–1022.

22

Granovetter, M. (1985): “Economic Action and Social Structure: the Problem of Embeddedness,” American Journal of Sociology, 91(3), 481– 493. (2005): “The Impact of Social Structure on Economic Outcomes,” Journal of Economic Perspectives, 19(4), 35–50. Jackson, M. (2009): “An Overview of Social Networks and Economic Applications,” forthcoming in the The Handbook of Social Economics, edited by J. Benhabib, A. Bisin, and M.O. Jackson, Elsevier Press. Jackson, M., and A. Wolinsky (1996): “A Strategic Model of Social and Economic Net- works,” Journal of Economic Theory, 71(1), 44–74. Krishnan, P., and E. Sciubba (2009): “Links and Architecture in Village Networks,” Economic Journal, 119(537), 917–949. Miguel, E., and M. Kremer (2003): “Networks, Social Learning, and Technology Adoption: The Case of Deworming Drugs in Kenya,” Mimeograph, University of California at Berkeley. Putnam, R. D. (1995): “Bowling Alone: America’s Declining Social Capital,” Journal of Democracy, 6(1), 65–78. Shipton, P. (1990): “How Gambians save and what their strategies imply for international aid,” World Bank Policy, Research and External Affairs Working Papers, WPS 395. Singh, N. (2000): “Theories of sharecropping,” in Readings in Development Economics, ed. by P. Bardhan, and C. Udry, vol. 1. MIT Press. Swindell, K. (1978): “Family Farms and Migrant Labour: The Strange Farmers of the Gambia,” Canadian Journal of African Studies / Revue Canadienne des tudes Africaines, 12(1), 3–17. Udry, C. (1996): “Gender, Agricultural Production, and the Theory of the Household,” Journal of Political Economy, 104(5), 1010–1046. Udry, C., and T. Conley (2004): “Social Networks in Ghana,” Yale University Eco- nomic Growth Center Discussion Paper n. 888. von Braun, J., and P. Webb (1989): “The Impact of New Crop Technology on the Agricultural Division of Labor in a West African Setting,” Economic Development and Cultural Change, 37(3), 513–534. 23

Webb, P. (1989): “Intrahousehold Decision-making and Resource Control: The Effects of Rice Commercialization in West Africa,” International Food Policy Research Center, Working Papers on Commercialization of Agriculture and Nutrition. Number 3.

24

Figure 1: INCOME DISTRIBUTION WITH DIFFERENT SURVEY INSTRUMENTS

Note: Kernel density estimates for the distribution of income with data for 15 villages in which both surveys were conducted. The sample for the household survey was randomly selected and collected disaggregated data on monthly income The distribution was built with 230 observations. The methodology for structured group interviews is described in the paper, the sample are all household heads and the distribution was built with 773 observations.

25

Figure 2: LAND MARKET NETWORK IN PASAMASSI FULA VILLAGE

26

Figure 3: LAND MARKET NETWORK IN PASAMASSI FULA VILLAGE

Note: This figure was produced using UCINET-NetDraw software.

27

Figure 4: OVERLAPPING NETWORKS IN PASAMASSI FULA VILLAGE

Labels: LAN D in red; LABOU R in blue; IN P U T in BLACK; CREDIT in green; Overlapping directed links in yellow. Note 1: Nodes are re-scaled according degree centrality. Note 2: This figure was produced using UCINET-NetDraw software.

28

29

Note: Village-level information. 60 observations for each variable.

Variable Approximated population Household size (between villages, median number of members) Household size difference (village standard deviation) Population density (persons/square kilometers) Income per capita (between villages, mean GMD) Gini (from self-declared income) Approximated available agricultural land (Hectares) Land per worker (between villages, mean hectares) Land per worker differences (village Standard Deviation, hectares) Diversity of ethnic groups (EDI) Diversity of main economic activity (village Herfindahl index) Diversity in educational level (village Herfindahl index) No formal education (village %) Koranic education (village %) No access to electricity (village %) No private toilettes (village %) Not improved water village %) Grasshuts (village %) Access to government news (% of TV and newspaper users) Marabout in the village

Mean Std. Dev. 585.74 245.44 11.57 3.28 9.91 9.42 6,901 4,118 3,208 1,705 0.34 0.11 337.66 363.79 2.69 3.53 4.65 6.08 0.30 0.23 0.47 0.17 0.29 0.12 82 10 61 18 97 4 38 30 88 20 38 29 16 16 0.5 0.5

Table 1: VILLAGE DESCRIPTION Min Max 202.00 1,402.00 7.00 21.50 3.35 61.18 652.21 20,310 450.65 11,695 0.13 0.66 0.00 2,044 0.00 23.82 0.00 27.93 0.84 0 0.81 0 0.53 0.05 48 100 7 100 84 100 0 100 0 100 0 94 0 57 0 1

Table 2: HOUSEHOLD DESCRIPTION Variable Age of household head Household Size Female Household head Compound head Polygamous Monogamous Non-Muslim Workers in the household (%) Agricultural land (hectares) Income per capita (GMD) Land per worker (hectares) Agricultural income (% of total) High quality of land Kamanyango Self respondent VDC member Officer above village level Relative of alkalo Member of elder council Traditional healer Emigrants

Mean 51.75 12.78 0.07 0.84 0.45 0.48 0.04 0.47 8.26 3,566 2.30 0.12 0.11 0.06 0.72 0.19 0.01 0.34 0.19 0.20 0.49

Std. Dev. 15.79 12.50 0.25 0.37 0.50 0.50 0.18 0.28 22.54 5,125 7.41 0.24 0.32 0.22 0.45 0.38 0.10 0.47 0.39 0.40 0.50

Min 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0

Max 100 400 1 1 1 1 1 1 400 125,000 133 1 1 1 1 1 1 1 1 1 1

Note: Household-level information. 3320 observations for each variable.

30

31

HOUSEHOLD LEVEL Degree External-in External-out (di (g m )) 0.019 0.082 0.055 0.038 0.274 0.228 0.022 0.068 0.040 0.251 0.028 0.078 0.031 0.044 0.268 0.174 0.028 0.738 0.600 0.042 0.440 0.490 0.016 0.122 0.048 0.033 0.327 0.214 0.087 0.091 29.2 16.6 29.9 20.8 39.8 22.5 49.9 40.9 23.0 17.9 147.7 76.5

Links

VILLAGE LEVEL Density Clusterness Compactness (Cl(g m )) (Cmp(g m )) 0.016 0.009 0.770 0.015 0.028 0.237 0.015 0.046 0.779 0.013 0.065 0.214 0.024 0.067 0.833 0.022 0.072 0.192 0.023 0.059 0.851 0.020 0.073 0.157 0.011 0.034 0.704 0.010 0.053 0.271 0.072 0.279 0.978 0.052 0.134 0.029

Note: 3320 observations for household-level information. 60 observations for village-level information.

Mean Std. Dev. LABOUR Mean Std. Dev. INPUTS Mean Std. Dev. MARRIAGE Mean Std. Dev. CREDIT Mean Std. Dev. KINSHIP Mean Std. Dev.

LAND

Network

Table 3: NETWORK DESCRIPTION

0.153 0.199 0.082 0.109 0.127 0.099 0.871 0.140 0.161 0.108 -

External Links

32

360 0.661

All -0.027*** (0.006) 0.002*** (0.001) 0.038** (0.017) -0.022** (0.010) -0.015 (0.011) -0.001 (0.002) 0.002 (0.001) 0.067* (0.038) 0.025*** (0.008) 0.0194 (0.012) 0.005 (0.010) -0.028 (0.025) 0.045*** (0.015)

(2) DENSITY Economic -0.015*** (0.004) 0.001** (0.001) 0.027* (0.014) -0.014* (0.008) -0.012 (0.009) -0.001 (0.001) 0.001 (0.001) 0.060* (0.035) 0.018*** (0.006) 0.004 (0.012) 0.011 (0.009) -0.033* (0.018) 0.016 (0.013) -0.008 (0.006) 240 0.598 120 0.708

Family -0.053*** (0.012) 0.003** (0.001) 0.058* (0.031) -0.039** (0.019) -0.020 (0.021) -0.002 (0.003) 0.003** (0.002) 0.089 (0.075) 0.041*** (0.014) 0.048* (0.024) -0.007 (0.020) -0.022 (0.049) 0.100*** (0.027)

(3)

(5) (6) CLUSTERNESS All Economic Family 0.009 0.023 -0.02 (0.016) (0.020) (0.038) 0.004* 0.005** 0.002 (0.002) (0.002) (0.006) -0.057 -0.009 -0.151 (0.037) (0.056) (0.108) 0.002 0.006 -0.004 (0.024) (0.032) (0.063) -0.098*** -0.073 -0.146 (0.034) (0.045) (0.098) 0.003 0.012** -0.014 (0.005) (0.005) (0.014) -0.003 -0.002 -0.005 (0.003) (0.005) (0.012) 0.102 0.135 0.034 (0.128) (0.153) (0.352) 0.024 0.0317 0.007 (0.019) (0.026) (0.037) -0.044 -0.03 -0.072 (0.046) (0.059) (0.103) 0.015 -0.008 0.063 (0.031) (0.044) (0.079) -0.118 -0.153* -0.048 (0.078) (0.090) (0.174) -0.027 -0.046 0.011 (0.036) (0.060) (0.117) 0.003 (0.026) 360 240 120 0.649 0.342 0.676

(4)

(8) (9) COMPACTNESS All Economic Family 0.013 0.029 -0.027 (0.043) (0.066) (0.029) 0.006 0.008 -0.003 (0.006) (0.008) (0.004) 0.389*** 0.477** 0.197** (0.126) (0.186) (0.082) -0.062 -0.053 -0.138* (0.101) (0.138) (0.071) -0.113 -0.101 -0.126* (0.107) (0.152) (0.072) 0.003 0.001 -0.012 (0.019) (0.027) (0.012) 0.009 0.008 0.010 (0.012) (0.016) (0.006) 0.997** 1.45** 0.265 (0.429) (0.658) (0.246) -0.033 -0.054 0.001 (0.049) (0.071) (0.048) 0.094 0.014 0.256*** (0.142) (0.212) (0.065) 0.089 0.216 -0.16*** (0.094) (0.143) (0.056) -0.143 -0.197 -0.121 (0.184) (0.255) (0.159) 0.125 0.060 0.209** (0.129) (0.179) (0.088) -0.222* (0.131) 360 240 120 0.385 0.417 0.609

(7)

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Clustered standard errors in parentheses. The variables Geographic density, Household size diversity, Education diversity, percentage of no formal education and the presence of a Marabout are included in the regression, but not reported for being always statistically insignificant.

Observations R-squared

External links

Semi-Urban

Female HH head

Grass Hut

No Water

No Private Toilettes

No Electricity

Land per worker

Income per capita

Activity Diversity

Ethnic Diversity

Gini

Household Size

Population

(1)

Table 4: DETERMINANTS OF NETWORK CHARACTERISTICS

Table 5: ECONOMIC AND DEMOGRAPHIC DETERMINANTS OF HOUSEHOLD CENTRALITY Xhh Household Size Age of Household Head Female Household Head Polygamous Household No-Muslim Formal Education Koranic Education Second quartile Third quartile Fourth quartile Compound Head Emigrants TV Observations R2

(1) LAND 0.007* (0.004) -0.001 (0.003) 0.000 (0.003) 0.001 (0.003) 0.002 (0.006) 0.002 (0.002) 0.001 (0.002) 0.002 (0.002) 0.000 (0.002) 0.001 (0.002) 0.005*** (0.002) 0.002 (0.001) -0.000 (0.002) 2784 0.281

(2) LABOUR 0.009** (0.004) 0.001 (0.003) 0.003 (0.003) -0.000 (0.003) 0.001 (0.006) 0.000 (0.002) 0.003* (0.002) 0.003* (0.002) 0.002 (0.002) 0.001 (0.002) -0.000 (0.002) -0.001 (0.001) 0.002 (0.002) 2784 0.334

(3) INPUTS 0.011*** (0.004) -0.004* (0.003) -0.001 (0.003) 0.005 (0.003) 0.011* (0.006) -0.003 (0.002) 0.003** (0.002) 0.004** (0.002) 0.005** (0.002) 0.003 (0.002) 0.001 (0.002) 0.000 (0.001) -0.000 (0.002) 2784 0.465

(4) CREDIT 0.004 (0.003) -0.000 (0.002) 0.000 (0.003) 0.001 (0.003) 0.000 (0.005) 0.002 (0.002) 0.003** (0.001) -0.001 (0.002) -0.000 (0.002) -0.003 (0.002) 0.000 (0.002) -0.000 (0.001) -0.003* (0.002) 2784 0.289

(5) MARRIAGE 0.004 (0.004) 0.004 (0.003) -0.000 (0.003) 0.006 (0.003) 0.001 (0.007) 0.002 (0.002) 0.001 (0.002) -0.002 (0.002) 0.001 (0.002) 0.001 (0.002) 0.010*** (0.002) 0.003* (0.002) -0.005** (0.002) 2784 0.605

(6) KINSHIP 0.023*** (0.007) 0.020*** (0.004) -0.011** (0.005) 0.007 (0.005) 0.004 (0.010) -0.006 (0.004) -0.002 (0.003) -0.004 (0.003) 0.003 (0.003) 0.004 (0.004) 0.009** (0.003) 0.000 (0.002) -0.002 (0.004) 2784 0.256

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors in parentheses Estimated by seemingly unrelated regressions (SUR) for the model presented in Equation (2). The variables Quality of land, Presence of Kamanyango, Land per worker, Percentage of active workers, Agriculture as proportion of the income, Proportion of newspapers readers and a dummy for selfrespondent interviewed were used in the estimation, together with the variables in Table 6 and 10, but not reported because of lack of statistic or economic relevance.

33

Table 6: HOUSEHOLD LEVEL CENTRALITY DETERMINANTS Xrole Alkalo Alkalo’s relative Alkalo’s assistant Elders Council head Elders Council member VDC Head VDC member Officer Griot Village Doctor Traditional Healer Marabout Imam Observations R2

(1) LAND 0.070*** (0.005) 0.002 (0.002) 0.006 (0.004) 0.001 (0.005) 0.001 (0.002) -0.001 (0.005) 0.003 (0.002) -0.005 (0.007) 0.008 (0.005) -0.002 (0.006) 0.003 (0.002) 0.003 (0.005) 0.008 (0.005) 2784 0.281

(2) LABOUR 0.016*** (0.005) 0.000 (0.002) 0.006 (0.004) 0.012** (0.005) -0.001 (0.002) 0.011** (0.005) 0.003* (0.002) -0.001 (0.007) 0.005 (0.005) 0.003 (0.006) 0.005*** (0.002) 0.006 (0.005) 0.017*** (0.005) 2784 0.334

(3) INPUTS 0.033*** (0.005) -0.000 (0.002) 0.011*** (0.004) -0.000 (0.005) 0.003 (0.002) 0.018*** (0.005) 0.004** (0.002) -0.002 (0.007) -0.007 (0.005) 0.010* (0.006) 0.004** (0.002) 0.003 (0.005) -0.004 (0.005) 2784 0.465

(4) CREDIT 0.012*** (0.004) -0.001 (0.001) 0.007** (0.003) 0.004 (0.005) -0.000 (0.002) 0.013*** (0.004) 0.001 (0.002) -0.000 (0.006) -0.003 (0.005) 0.005 (0.005) 0.001 (0.001) 0.006 (0.004) -0.008* (0.004) 2784 0.289

(5) MARRIAGE -0.003 (0.005) -0.001 (0.002) 0.010** (0.004) 0.014** (0.006) 0.004 (0.003) -0.002 (0.006) 0.003 (0.002) -0.005 (0.007) 0.003 (0.006) 0.016** (0.010) 0.000 (0.002) 0.005 (0.005) 0.007 (0.005) 2784 0.605

(6) KINSHIP 0.094*** (0.008) 0.020*** (0.003) 0.026*** (0.007) 0.028*** (0.009) -0.001 (0.002) 0.046*** (0.009) 0.009*** (0.003) 0.010 (0.012) -0.008 (0.009) 0.027*** (0.007) 0.008** (0.003) 0.007 (0.008) 0.002 (0.009) 2784 0.256

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors in parentheses Estimated by seemingly unrelated regressions (SUR) for the model presented in Equation (2). Variables in Table 5 and 10 are also included in the estimation.

34

Table 7: DYADIC REGRESSION FOR DETERMINANTS OF LINK FORMATION Xhh Same ethnic group Sum of land Difference in land Both with productive land Sum % agricultural income Difference % agric. income Just one female Just one with formal educ. Both with formal education Just one with koranic educ. Both with koranic education Just one have emigrants Both have emigrants Both with kamanyango Sum of the age Difference of age Sum income per capita Difference inc. per capita Sum household size Difference household size Sum of workers Observations

(1) LAND -0.024 (0.085) 0.152*** (0.022) 0.198*** (0.024) 0.501** (0.201) 0.197 (0.146) 0.078 (0.176) 0.010 (0.129) -0.186** (0.088) -0.237 (0.237) 0.063 (0.078) 0.006 (0.120) -0.043 (0.090) 0.079 (0.108) 0.339 (0.259) -0.103 (0.095) -0.172 (0.123) 0.013 (0.020) -0.010 (0.020) 0.228*** (0.061) 0.105* (0.059) 0.206 (0.144) 68198

(2) LABOUR 0.590*** (0.087) 0.021*** (0.008) 0.005 (0.009) 0.577*** (0.159) 0.509*** (0.124) -0.220 (0.164) 0.133 (0.129) -0.135 (0.090) -0.015 (0.209) 0.104 (0.077) 0.159 (0.109) -0.124* (0.075) -0.159 (0.098) 0.693*** (0.253) 0.064 (0.090) 0.056 (0.113) 0.017 (0.017) -0.007 (0.017) 0.332*** (0.059) 0.128* (0.077) 0.289** (0.138) 68198

(3) INPUTS 0.386*** (0.083) 0.023*** (0.006) 0.001 (0.007) 0.494*** (0.154) 0.191* (0.109) -0.068 (0.137) -0.383*** (0.133) -0.202*** (0.075) -0.201 (0.191) 0.136** (0.065) 0.288*** (0.090) -0.090 (0.067) -0.142* (0.084) -0.125 (0.261) -0.135* (0.071) -0.175* (0.103) 0.015 (0.015) -0.011 (0.015) 0.290*** (0.040) -0.055 (0.053) 0.282** (0.120) 68198

(4) CREDIT 0.129 (0.116) 0.006 (0.011) 0.001 (0.010) 0.612** (0.285) 0.070 (0.203) 0.668*** (0.222) 0.012 (0.144) 0.351*** (0.136) 0.957*** (0.246) 0.424*** (0.120) 0.569*** (0.192) -0.244** (0.112) -0.328* (0.191) 0.376 (0.381) 0.134 (0.124) -0.193 (0.127) 0.059*** (0.014) -0.039*** (0.014) 0.412*** (0.096) 0.178** (0.084) 0.030 (0.288) 68198

(5) MARRIAGE 1.021*** (0.091) 0.033*** (0.007) -0.012 (0.009) -0.221 (0.149) 0.088 (0.120) -0.015 (0.145) -0.072 (0.095) -0.157** (0.068) -0.100 (0.182) 0.025 (0.062) 0.112 (0.087) 0.020 (0.071) 0.179** (0.081) 0.191 (0.230) 0.340*** (0.082) 0.240** (0.094) 0.016 (0.012) -0.020 (0.014) 0.665*** (0.039) -0.033 (0.051) 0.490*** (0.105) 68198

(6) KINSHIP 1.372*** (0.073) 0.026*** (0.004) -0.032*** (0.005) 0.307*** (0.098) 0.323*** (0.084) -0.266** (0.111) -0.218*** (0.059) -0.087* (0.045) -0.055 (0.101) -0.123*** (0.041) 0.105* (0.059) -0.090** (0.042) -0.066 (0.054) 0.355* (0.185) 0.295*** (0.051) -0.155** (0.071) 0.012* (0.007) -0.015* (0.008) 0.218*** (0.028) 0.021 (0.032) 0.185** (0.076) 68198

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Two dimensions (i and j) clustered standard errors in parentheses. Logit estimates for the model presented in equation (3). Village’s dummies were included but are not reported.

35

Table 8: DYADIC REGRESSION FOR DETERMINANTS OF LINK FORMATION Xrole Alkalo VDC head Just one is VDC member Both are VDC members Just one is Alkalo’s assistance Just one is Alkalo’s relative Both are Alkalo’s relative Just one from Elders Council Both from Elders Council Imam Just one compound head Both are compound head Observations

(1) LAND 1.174*** (0.134) 0.032 (0.164) 0.056 (0.077) 0.548*** (0.155) 0.298** (0.138) -0.002 (0.094) 0.175 (0.135) 0.092 (0.087) 0.496*** (0.158) 0.316* (0.167) 0.858*** (0.257) 1.219*** (0.261) 68198

(2) LABOUR 0.518*** (0.140) 0.424*** (0.126) 0.159** (0.064) 0.408*** (0.129) 0.205* (0.115) -0.238*** (0.091) -0.067 (0.135) 0.055 (0.073) 0.039 (0.152) 0.521*** (0.199) -0.387** (0.154) -0.262 (0.160) 68198

(3) INPUTS 0.560*** (0.096) 0.405*** (0.117) 0.105* (0.059) 0.391*** (0.112) 0.291** (0.119) -0.137* (0.074) -0.058 (0.105) 0.149** (0.062) 0.161 (0.145) -0.011 (0.127) -0.251* (0.149) -0.017 (0.153) 68198

(4) CREDIT 0.854*** (0.199) 0.410* (0.236) 0.271** (0.115) 0.662*** (0.192) -0.129 (0.185) 0.142 (0.132) 0.094 (0.196) 0.263** (0.130) 0.593*** (0.194) -0.584*** (0.186) 0.292 (0.221) 0.462* (0.278) 68198

(5) MARRIAGE 0.302*** (0.110) 0.102 (0.140) 0.100* (0.054) 0.187 (0.132) 0.161* (0.098) -0.107* (0.065) -0.140 (0.095) 0.111* (0.061) 0.238** (0.119) 0.138 (0.109) 0.622*** (0.195) 1.182*** (0.194) 68198

(6) KINSHIP 0.773*** (0.080) 0.483*** (0.090) 0.126*** (0.035) 0.257*** (0.074) 0.214*** (0.072) -0.058 (0.052) 0.316*** (0.077) 0.062 (0.043) 0.219*** (0.081) 0.126 (0.080) 0.002 (0.080) 0.180** (0.087) 68198

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Two dimensions (i and j) clustered standard errors in parentheses. Logit estimates for the model presented in equation (3). Village’s dummies were included but are not reported.

Table 9: SIMPLE CORRELATION IN DEGREE LAND LAND 1 LABOUR 0.2879 INPUTS 0.2603 MARRIAGE 0.0532 KINSHIP 0.2336 0.128 CREDIT

LABOUR

INPUTS

MARRIAGE

KINSHIP

CREDIT

1 0.4608 0.1006 0.3553 0.1889

1 0.1405 0.3282 0.2554

1 0.2421 0.2046

1 0.1777

1

36

Table 10: HOUSEHOLD LEVEL CENTRALITY DETERMINANTS

dextin,land dextout,land dextlabour dextin,inputs dextout,inputs dextin,credit dextout,credit dextin,marriage dextout,marriage Observations R2

(1) LAND -0.010*** (0.003) -0.006* (0.003) 0.010*** (0.003) 0.008*** (0.003) -0.007* (0.004) 0.003 (0.002) 0.008** (0.003) -0.001 (0.002) -0.001 (0.002) 2784 0.281

(2) LABOUR 0.001 (0.003) 0.003 (0.003) -0.003 (0.003) -0.000 (0.003) -0.003 (0.004) 0.000 (0.002) 0.009*** (0.003) -0.002 (0.002) 0.001 (0.002) 2784 0.334

(3) INPUTS -0.002 (0.003) 0.002 (0.003) 0.004 (0.003) -0.009*** (0.003) -0.003 (0.004) 0.001 (0.002) -0.001 (0.003) -0.001 (0.002) 0.003* (0.002) 2784 0.465

(4) CREDIT 0.002 (0.002) 0.003 (0.003) -0.004* (0.002) 0.001 (0.002) 0.005 (0.003) -0.005*** (0.002) 0.000 (0.003) -0.002 (0.001) 0.001 (0.001) 2784 0.289

(5) MARRIAGE -0.006** (0.003) 0.003 (0.004) 0.005* (0.003) -0.000 (0.003) 0.011*** (0.004) -0.002 (0.002) -0.003 (0.003) -0.003* (0.002) -0.002 (0.002) 2784 0.605

(6) KINSHIP 0.004 (0.005) 0.009* (0.006) 0.004 (0.005) -0.007 (0.004) 0.006 (0.007) -0.002 (0.004) 0.005 (0.005) -0.004 (0.003) 0.001 (0.003) 2784 0.256

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Standard errors in parentheses Estimated by seemingly unrelated regressions (SUR) for the model presented in Equation (2). Variables in Table 5 and 6 are also included in the estimation.

Table 11: DEGREE ACROSS NETWORKS VARIABLES

LAND

LAND LABOUR INPUT CREDIT MARRIAGE KINSHIP

0.258*** (0.019) 0.106*** (0.020) 0.030 (0.022) -0.003 (0.018) -0.004 (0.011)

LABOUR

INPUT

CREDIT

MARRIAGE

0.238*** (0.018)

0.097*** (0.018) 0.486*** (0.017)

0.021 (0.016) 0.008 (0.017) 0.097*** (0.017)

-0.003 (0.020) -0.040* (0.021) 0.121*** (0.021) 0.456*** (0.023)

0.487*** (0.018) 0.010 (0.021) -0.033* (0.017) 0.034*** (0.010)

37

0.126*** (0.021) 0.099*** (0.017) 0.039*** (0.010)

0.288*** (0.014) 0.009 (0.009)

0.066*** (0.012)

Table 12: ACTOR CORRELATIONS AMONG NETWORKS obs mean LAND LABOUR 453 0.13 INPUTS 564 0.15 MARRIAGE 490 0.05 835 0.04 KINSHIP CREDIT 277 0.09 LABOUR INPUTS 641 0.31 MARRIAGE 515 0.10 886 0.19 KINSHIP 341 0.18 CREDIT INPUTS MARRIAGE 647 0.08 KINSHIP 1141 0.18 CREDIT 399 0.17 MARRIAGE KINSHIP 1100 0.24 CREDIT 319 0.08 KINSHIP CREDIT 554 0.15

38

std. dev. 0.31 0.32 0.21 0.20 0.27 0.38 0.29 0.29 0.33 0.26 0.28 0.32 0.31 0.24 0.27

median -0.02 -0.02 -0.03 -0.03 -0.02 0.20 -0.02 -0.01 -0.02 -0.03 -0.01 -0.02 0.21 -0.02 -0.02

min max -0.17 1 -0.20 1 -0.21 1 -0.36 1 -0.19 1 -0.21 1 -0.32 1 -0.25 1 -0.13 1 -0.20 1 -0.24 1 -0.21 1 -0.22 1 -0.18 1 -0.18 1

Table 13: DYADIC REGRESSION FOR EXTERNAL LINKS

Just one dextin,land Just one dextout,land Both dextin,land Both dextout,land Just one dextlabour Both dextlabour Just one dextin,inputs Just one dextout,inputs Both dextin,inputs Both dextout,inputs Just one dextin,credit Just one dextout,credit Both dextin,credit Both dextout,credit Just one dextin,marriage Just one dextout,marriage Both dextin,marriage Both dextout,marriage Observations

(1) LAND -0.618*** (0.143) -0.108 (0.129) -1.516*** (0.427) -0.787** (0.328) 0.073 (0.112) 0.712*** (0.199) 0.196** (0.093) -0.255* (0.139) 0.131 (0.285) 0.000 (0.000) 0.057 (0.082) 0.157 (0.104) 0.317 (0.222) 0.603 (0.399) -0.137 (0.150) 0.240** (0.114) -0.155 (0.164) 0.188 (0.136) 68198

(2) LABOUR 0.071 (0.107) -0.009 (0.137) 0.700** (0.281) 0.195 (0.225) -0.020 (0.099) -0.170 (0.234) -0.049 (0.096) -0.006 (0.120) 0.181 (0.221) -0.807 (0.736) 0.058 (0.080) 0.178* (0.104) -0.141 (0.217) -0.225 (0.454) -0.274** (0.119) 0.190* (0.104) -0.174 (0.127) 0.258** (0.117) 68198

(3) INPUTS -0.138 (0.096) 0.111 (0.100) 0.376 (0.241) 0.037 (0.229) 0.149* (0.081) 0.433** (0.213) -0.217** (0.084) -0.085 (0.109) -0.213 (0.188) -1.303* (0.719) -0.020 (0.070) 0.065 (0.085) -0.054 (0.186) -1.019* (0.571) 0.060 (0.118) 0.090 (0.085) 0.088 (0.124) 0.176* (0.093) 68198

(4) CREDIT 0.100 (0.204) 0.098 (0.219) 0.840*** (0.269) -0.108 (0.487) 0.259 (0.160) 0.551* (0.324) -0.016 (0.179) 0.241 (0.208) 0.029 (0.262) -0.846 (0.516) -0.184 (0.115) -0.318** (0.158) -0.068 (0.287) -0.628 (0.458) -0.200 (0.139) -0.107 (0.104) -0.150 (0.205) -0.253 (0.170) 68198

(5) MARRIAGE -0.089 (0.099) 0.081 (0.099) -0.167 (0.186) 0.295 (0.212) -0.035 (0.095) -0.106 (0.255) -0.056 (0.080) 0.070 (0.116) -0.238 (0.201) -0.116 (0.516) -0.113* (0.067) 0.030 (0.099) 0.046 (0.173) -0.539 (0.417) -0.189** (0.086) 0.050 (0.084) -0.731*** (0.095) 0.123 (0.091) 68198

(6) KINSHIP 0.010 (0.061) 0.020 (0.070) 0.268* (0.163) 0.178 (0.139) 0.016 (0.062) -0.069 (0.176) -0.012 (0.052) -0.057 (0.069) 0.036 (0.132) -0.028 (0.313) -0.024 (0.043) 0.128* (0.066) -0.039 (0.122) 0.503** (0.256) -0.137** (0.062) 0.005 (0.052) -0.125* (0.067) 0.059 (0.060) 68198

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Two dimensions (i and j) clustered standard errors in parentheses. Logit estimates for the model presented in equation (3). Village’s dummies were included but are not reported. Same controls as in Table 7 and Table 8.

39

Table 14: DYADIC REGRESSION FOR EXTERNAL LINKS (1) LAND LAND LABOUR INPUTS CREDIT MARRIAGE KINSHIP Observations

0.065*** (0.011) 0.055*** (0.007) 0.010** (0.005) 0.020*** (0.007) 0.003 (0.004) 68198

(2) LABOUR 0.067*** (0.010)

0.170*** (0.017) 0.027*** (0.007) 0.029*** (0.009) 0.050*** (0.006) 68198

(3) INPUTS 0.075*** (0.011) 0.226*** (0.019)

0.044*** (0.009) 0.022** (0.010) 0.064*** (0.008) 68198

(4) CREDIT 0.024** (0.010) 0.061*** (0.012) 0.075*** (0.015)

0.036*** (0.010) 0.033*** (0.006) 68198

Note: *** p < 0.01, ** p < 0.05, * p < 0.1. Two dimensions (i and j) clustered standard errors in parentheses. Logit estimates for the model presented in equation (3). Village’s dummies were included but are not reported. Same controls as in Table 7, Table 8 and Table 13.

40