The Effect of Ethnic Heterogeneity on the Formation of Informal Insurance Networks in South India

Transcription

1 Haverford College Economics Senior Thesis The Effect of Ethnic Heterogeneity on the Formation of Informal Insurance Networks in South India Mallimalika Gupta Shannon Mudd, Advisor April 26, 2012 Abstract This paper looks at the effect that two different indices designed to summarize ethnic heterogeneity have on the formation of informal insurance networks in 75 villages in south India. Using extensive network data, a measure for the borrowing and lending behavior is constructed. An OLS framework is used to observe the impact that ethnic polarization and fractionalization have on borrowing and lending of money and of rice and kerosene in these communities. The paper finds an inverse, non-monotonic relationship between the measures of ethnic heterogeneity and strength of the lending and borrowing measures that reverses at a certain threshold. This shows that the size, number, and relative positional power of ethnic groups in a community affect households decision to lend and borrow from others in the community. Homogenous and moderate- to highly heterogeneous communities are best able to pool risk.

2 ACKNOWLEDGEMENTS I would like to express my gratitude to my advisor Dr. Shannon Mudd for constant encouragement, guidance and support throughout writing this thesis. I would like to thank Dr. Paul Cichello for continuous help on several econometric and other issues during the entire process. Also, I am grateful to Dr. Anne Preston, Dr. Saleha Jilani and Dr. Indradeep Ghosh for their comments. I am deeply appreciative of my undergraduate classes in the Tri-Colleges for teaching me to be critical, rigorous and creative at the same time. Truly, this thesis is a culmination of ideas that emerged from classes across the disciplines. Last, I want to thank my family. 2

3 Contents Introduction... 4 Literature Review... 7 Setting Data Theoretical framework Empirical Model Results Discussion Conclusion Bibliography Tables Appendix A Setting: Ethnic background in India and a look at Karnataka Appendix B - Choice of Empirical Method

4 Introduction Risk management is integral to the economic advancement and survival of people in lowincome, agrarian countries with uncertain streams of income. An informal insurance network helps smooth consumption in the face of unexpected shocks to income and in the absence of other formal sources of income. A number of different determinants can affect the formation of such networks. Previous research has identified an individual s demographic characteristics, popularity, education status, class, and other community-level characteristics as factors that have an effect on the formation of connections between individuals in a community. This paper investigates the effect of a community-level characteristic the extent of ethnic heterogeneity along religious and caste-lines on the formation of informal insurance networks in that community. Increasing heterogeneity in racial and ethnic groups has been shown to negatively affect community welfare and public goods spending. This paper ties together two different lines of research to look at the effect of the fractionalization 1 of both religion and caste in a village on the extent to which a household is willing to borrow or lend either money or rice and kerosene from other households in the village. The paper uses extensive network data and demographic information collected in seventy-five villages in south India to answer this question. An alternative measure for ethnic diversity, the polarization index, is also used in the analysis. The paper finds a significant non-linear relationship between both religious fragmentation and polarization and the lending and borrowing behavior of households in the village. For caste, we observe that the polarization index produces statistically significant results that are consistent for all measures of borrowing and lending while the fragmentation index produces a significant 1 This paper uses the words fragmentation, fractionalization and ELF (ethnolinguistic fractionalization index) interchangeably. They all stand for the fractionalization index. 4

5 impact on the borrowing and lending of rice and kerosene. We also observe that all of these relationships are non-linear, approaching turning points where the inverse relationship reverses. Thus, highly homogenous and highly heterogeneous communities are best able to pool risk. The results suggest that the size and number of groups in a multi-ethnic setting makes a significant impact on the formation of informal insurance networks in the community. Previous studies have identified a negative impact of ethnic heterogeneity on various economic and political economy variables. The literature on public goods provision finds that a higher degree of fractionalization leads to a lower level of public goods provision, and implicitly in some cases, of trust. Indirectly, this often translates to lower economic growth. The persistent difficulties Sub-Saharan Africa has faced with respect to achieving economic growth, for example, have been blamed on the linguistic and ethnic fragmentation of this region (Easterly and Levine, 1997). The basic hypothesis is that with more ethnic groups (based on language, religion, tribe, etc.), communication becomes harder, trust reduces and political instability rises, the ease of doing business decreases, and economic growth is adversely affected. This logic may be well applied to India as well. India is extremely diverse, with over 2000 ethnic groups. Every major religion in the world is represented. When India gained independence from colonial Great Britain, one of the main dilemmas facing the first Prime Minister Nehru was how to build the state given the history of conflict between religious groups and the potential of conflict arising from linguistic and ethnic differences. However, over the last sixty-five years of independence, India has been largely peaceful, and in recent years its economy has been growing at an average of 7.5% a year, a very different outcome from sub-saharan Africa and its failure to grow. Because of this reality of the situation, the study of the effect of ethnic diversity on the formation of informal insurance networks in this setting becomes even more interesting. 5

6 Do people in a community feel more comfortable asking others in their community for help if more people look the same in terms of the caste they belong to, and the religion they follow? Or, are they more likely to ask for help in the case where there is a greater diversity of ethnic and caste groups? The paper uses the fractionalization (ELF) index, calculated using the Herfindahl formula (Labar, 2010), and the polarization index of Montalvo and Reynal-Querol (2003), in order to answer this question. The ELF index denotes the probability that two persons chosen at random in a community are from two different ethnic groups. A high ELF index indicates a larger degree of heterogeneity within the community. The index increases with the number of ethnic groups in a community and as the size of the groups become more equal. The polarization index is an alternative indicator of ethnic diversity. This index reaches a maximum when the groups are of equal size. It captures both the number of groups and the threat that one group might feel because of the increased size of another. For a given number of groups, the threat is higher the larger the size of another group relative to the size of the reference group (Montalvo and Reynal-Querol, 2003). An empirical study of social networks is significant because they provide mutual insurance in developing countries. Such networks facilitate active interactions amongst individuals in a community, leading to transmission of knowledge, increases in aggregate human capital, and the development of 'trust', which improves the functioning of markets (Alesina and Le Ferrara, 1999). Social networks make communities more closely-knit, developing higher sense of trust in the community. Trust in strangers facilitates collective action. Networks of personal contacts are important in obtaining information about job opportunities, in optimal provision of public goods, and formation of alliances, cartels, and federations. Among poor people, small transfers can have large welfare impacts if that transfer is able to move the recipient to a higher expected welfare level (Santos and Barrett, 2010). Social networks have long been identified as the key loci of 6

7 informal risk sharing (Vanderpuye-Orgle and Barrett, 2009). There are two different types of risk idiosyncratic risk such as a death in the family, which only concerns one person or one household at a time, and covariate risk, such as a natural disaster, which affects an entire community. The focus of this paper is on idiosyncratic risk management. The unexpected shocks to income that would lead a person to ask for help are inherently idiosyncratic in nature. In the case of a covariate risk, the shock to income is universal, so informal insurance networks within a community will not be effective in consumption smoothing. This paper looks closely at the effect that the social fabric of a community can have on individual decision-making. The results have implications for other parts of India, and beyond. Questions of ethnic heterogeneity and its effect on community-building, public goods allocation, voter preferences, etc. are relevant in all societies. Moreover, the results of this paper have implications for other areas such as lending methodologies in microfinance. It sheds light on the type of communities that would foster group-lending methodologies versus individual methodologies, for instance. Literature Review Previous research related to this question falls under two different strands. One strand of literature looks specifically at the determinants of social and informal insurance network formation, credit sharing and risk sharing, and the other looks at the effect of ethnic heterogeneity on the provision of public goods, political stability and other economic indicators. The literature on ethnic heterogeneity uses the fractionalization index in empirical investigation of questions that deal mainly with the quality of institutions and public goods provision. Two 7

8 political economy theories help ground our specific question. The literature review will detail these theories along with papers that found results significant to this study of informal insurance networks and ethnic heterogeneity. Past Empirical Evidence Previous research has found significant (mainly inverse) impact of ethnic fragmentation on the allocation of public goods, economic growth, and participation in social activities. Also, a significant impact of ethnic polarization on the onset of conflict (Montalvo and Reynal-Querol, 2007) has been observed. The effect of ethnic fragmentation on growth and other economic indicators has been studied extensively, starting with Easterly and Levine s (1997) popular paper, Africa s Growth Tragedy. This paper hypothesizes and shows that ethnic diversity explains differences in public policies, political stability and other economic indicators, which directly affect the growth of a country. This strand of literature makes use of political economy models that state that polarized societies will be prone to competitive rent-seeking behavior, which makes it difficult to come to a consensus regarding public goods like infrastructure, education and policies. After accounting for the effects of ethnic diversity on education, political stability, financial depth, black market premia, fiscal policy, and infrastructure development, Easterly and Levine find that ethnic diversity alone accounts for about 28 percent of the growth differential between the countries of Africa and East Asia. The paper concludes that the high level of ethnic diversity in Africa explains its tragic growth performance. Community behavior with respect to provision of public goods is inversely affected by ethnic fractionalization (Alesina, Baqir and Easterly, 1999, Poterba 1996). While voting for public welfare, individuals seem to demonstrate racial group loyalty (Luttmer, 2001). 8

9 Individuals increase their support for welfare spending as the share of local recipients from their own racial group rises. Interpersonal preferences for welfare can be used to explain redistribution of wealth. Luttmer finds that within the United States, relatively racially heterogeneous states provide lower welfare benefits. The self-reported preference for income distribution that Luttmer investigates is driven by similar tendencies that lead to the self-reported preferences for borrowing and lending patterns being investigated in this paper. Chen (2007) uses the fractionalization and polarization measures to test cross-group attitudes and preferences for a public good education within communities in California. Chen captures the effect of ethnic heterogeneity on preferences for higher public education spending, using an OLS regression framework with the percentage of yes votes for Proposition 39 (also known as the Smaller Classes, Safer Schools and Financial Accountability Act ) as the dependent variable. The paper finds a U-shaped relationship between racial fragmentation and preferences for public education spending, ceteris paribus. Preferences for public school spending are higher in communities with extremely low and high levels of racial fractionalization, and are lowest in communities with intermediate levels of fractionalization. This paper advances the possibility of a non-monotonic relationship between community level composition of ethnic groups and preferences for a public good. Ethnic fragmentation also reduces the participation in social activities. Alesina and La Ferrara (1999) find that participation in social activities declines and becomes unequal in more racially or ethnically fragmented localities in the US. In keeping with Putnam s (1995) theory that the strength of civil society is positively related to the participation of its members in social activities like religious groups, hobby clubs, youth groups, sports groups, etc., this paper looks at the role of heterogeneity in income, race and ethnicity in dictating the strength and creation of 9

10 such social interaction. Varshney (2002) studies the relationship from a different angle by observing the impact of strong civil society (which denotes an idea of social capital, similar to Putnam) on the breakout of religious conflict in India. He finds that cities with stronger associational networks 2 are less likely to experience riots or religious conflict between Hindus and Muslims. Overall, research has found that ethnic diversity and the ethnic composition of the community is indicative of the level of participation in activities of a communal nature and the level of trust within the community. Most studies find a negative relationship between racial and ethnic heterogeneity and provision of public goods or formation of communal networks. A number of researchers study the observable demographic or other characteristics of a group and the extent to which these variables affect the ties formed within a network. These studies find that characteristics encompassing identity, ethnicity, and groups have a significant impact on the formation of networks. DeWeerdt (2004) finds that kinship, geographical proximity, number of common friends, clan membership, religious affiliation and wealth strongly determine the formation of risk-sharing networks in rural Tanzania. He also finds that poor households have less dense networks than the rich. His dependent variable measures the likelihood of forging a relationship or an insurance link across all possible pairs in a village. Santos and Barrett (2010) study the determinants of bilateral credit sharing by looking at a person s willingness to give a loan of cattle to another person in a pastoral community in southern Ethiopia. They construct a dyadic (pair-wise) probit model that tests the likelihood of a link being formed based on wealth and other physical and socioeconomic covariates. They find a 2 Associational networks are civil, non-governmental associations such as bowling leagues, which is Putnam s poster example in his popular book Bowling Alone: The Collapse and Revival of American Community (2000). He blames the decline in America s social capital on the decline in the memberships of bowling leagues and other such associational groups. 10

11 middle class bias, with both the poor and the rich being less likely to able to collect credit. They also find that the propensity to lend cattle is strongly and positively influenced by the clan identity. This paper adds to the argument that demographic and wealth characteristics of an individual affect the formation of their social networks as well. Vanderpuye-Orgle and Barrett (2009) find that social visibility, which is based on certain demographic characteristics of an individual, affects the extent of informal insurance available to individuals in rural Ghana. The socially invisible, who tend to be younger, engaged in farming, recent arrivals to the village who have been fostered and are not members of a major clan, are unable to pool risk. Network formation has also been studied using games. Using a sender-receiver investment game played by Harvard students, Glaeser et al (2000) find that senders send lower amounts to receivers when the receiver s race is different from their own. They observe that those with a higher social status (parents education, proxies for wealth, volunteer organization membership) are returned larger amounts by the receiver. Breza, Chandrashekhar and Larreguy (2011) create a field experiment, setting up a similar two party sender-receiver trust game, using the same dataset as the one used for this project 3. They find that both senders and receivers in socially close pairs make larger transfers to each other. Leaders in the community receive higher transfers, as do high caste individuals in the community. Overall, there is strong evidence for borrowing and lending patterns in informal networks being influenced by perceived and visible social and ethnic characteristics. 3 Other work done using the same data include Banerjee et al (2011), who observe the diffusion of the decision to pursue microfinance through social networks and find that participants are more likely to pass on information to friends and acquaintances than informed non-participants. Jackson, Rodriquez-Barraquer and Tan (2011) look at two measures of social network formation clustering and support within this data. Clustering measures the extent to which two friends of a given agent are friends of each other. They find that in this data clustering is on the order of ten to thirty percent. The concept of support measures the number of pairs of friends that have some other friend in common. Support is several times higher than clustering here. 11

12 Power Threat and Contact Theory Theoretically, two main theories from political economy and social psychology look at how a difference in the composition of groups affects the way people behave. These two theories are predicated by the group position model (Blumer, 1958). The model predicts that inter-group racial prejudices are affected by a sense of group position relative to others in the community. This consideration of the positional arrangement of ethnic or racial groups with respect to other groups can be applied to relations in multiethnic settings like the villages in this study. Chen (2007) points out that this model assumes that the dominant ethnic group must maintain a crucial distinction between themselves and subordinate group members. The distinction includes perceptions of abilities, characteristics, and expected behaviors of the minority group members. This dynamic of group relations is driven by perceptions of threat that groups have about each other. Two hypotheses look at these group dynamics in a community. Chen develops the hypotheses from studying other work in the area, and here Chen s depiction of the theories is recounted. The first is the power-threat hypothesis which says that any animosity that the dominant group may feel for the subordinate group will increase with the size of the minority group. The dominant group anticipates an increased threat to its economic and social status as the minority group increases in number of members. Oliver and Wong (2003) predict that this theory best explains relationships between two groups that have had a history of tense relations in the past 4. The second hypothesis is the contact hypothesis, which predicts that a larger proportion of out-group (non-dominant) group members in a community may improve intergroup relations. This theory stems from the assumption that individuals feel more comfortable around people that are most like themselves and may be less tolerant of those who are different. Increased number 4 In India, this would be Hindus and Muslims, with respect to religion, and upper castes versus schedules castes and scheduled tribe groups, with respect to castes, as will be explained in the Setting section and in the Appendix. 12

13 of people from the non-dominant group will lead to increased exposure across groups, which will help correct negative ethnic stereotypes. The contact theory and the power-threat hypothesis stand in opposition to each other, because it seems that the contact hypothesis predicts a positive effect on the inter-group relations with an increase in the number and size of groups, and the power-threat hypothesis seems to suggest the opposite. Chen (2007) develops a setting whereby she combines these two explanations to suggest a non-monotonic relationship between racial fractionalization and interracial relations in a community, and in turn, preferences for a public good in that community. This paper uses the fractionalization and the polarization measures, because both help capture the two theories, and each seems to explain one theory better than the other. The polarization index seems to better explain the power threat hypothesis, because in this type of index, what matters is not only how many groups there are, but also if they view other groups as a potential threat for their interests (Montalvo and Reynal-Querol, 2003). The fractionalization index observes the number and size of other groups that help explain the contact theory. The Data and Theoretical Framework sections, further help explain why incorporating both indices in our analysis is valuable. Setting Our study is set in a country that presents a very interesting landscape with respect to ethnic diversity. India is second only to the continent of Africa with respect to its linguistic, genetic and cultural diversity. The concentration of such large numbers of ethnically diverse people within the subcontinent has not always been peaceful. A large part of modern India s history has been marked by religious and caste-based conflict. The two main religions in India with respect to 13

14 religious conflict are Hinduism and Islam. Followers of the two religions have a long history of tension that continued during colonial times and into the period after Independence. The secular democratic country also contains large number of Christians and is home to the world s largest number of Zoroastrians, Sikhs, Hindus, Jains and Bahai s. Caste is a form of division attached to socioeconomic status and is transferred down a family. One is born to a caste. Originally, it was used an indicator to enforce a division of labor; caste was tied to profession. Hence, it became tied to economic opportunity and soon to privilege and standard of living. There are three groups of which to be cognizant: the upper castes, or the Brahmins, Kshatriyas and Vaishyas; the exuntouchables or dalits also known as the Scheduled Castes (SCs); and the middle category, consisting of the peasants, artisans, etc, who are known as the OBCs, or Other Backward Castes. The SCs were the most discriminated against in the past. The OBCs had numbers on their side, and were not as strongly discriminated against as the SCs. Currently, a number of reservations exist in legislative bodies and in government-run educational institutions as well as offices for people belonging to the groups that were discriminated against in the past. For a more detailed description of the history of the religious tensions, and background of this study, please see Appendix A. Data To investigate the effect of heterogeneity on the formation of informal insurance networks, this paper utilizes a unique dataset of survey information on 75 villages in south India 5. The dataset consists of network and demographic information of the residents of the villages, all of which are located in Karnataka, outside Bangalore (the capital of the state). 5 The data was collected by a team that is part of the Jameel Poverty Action Lab (JPAL). JPAL originally collected the data to study the diffusion of microfinance in these villages. These 75 villages had been identified by Bharatha Swamukti Samsthe, a microfinance institution, as a place where it planned to begin operation. 14

15 Survey The research team collected information from the villages in two different efforts, one at the household level and one at the individual level. First, basic information was collected on all households, i.e., a full census with demographic information. This household level data includes information on the caste and religion of the members. In addition, it contains information that can be used to construct variables indicative of relative wealth of the households for instance, whether the house has electricity, the number of rooms and beds in the house, the type of roof of the house, and whether the house is rented or owned by the occupant. This information is present for the entire population of each village. In addition to the full village census of the households, a second dataset was collected on a subsample of individuals in each village. This dataset contains more detailed information, such as language, occupation, gender, age, education, etc. The sub-sample was stratified by religion and by geographic location. About 46% of the population was included in these samples. On average, each village has 216 households. The largest village has 356 households, and the smallest village has 77 houses. Each village has a Hindu majority, which is common in India. Some villages are entirely Hindu, but others have Muslims and Christians. The village with the lowest Hindu majority in the dataset have 56.86% Hindus. The data groups caste into five groups General, OBC, SC (Scheduled Caste), ST (Scheduled Tribe) and Minority. The largest share of households (37.02%) belongs to the OBC (Other Backward Classes) group. Only 9.2% of the population belongs to the general category, i.e. the castes groups that were originally considered upper class. A majority (89%) of the houses were owned by the people that lived in them. The rest of the households (5.09%) were on rent, while a few families (4.13%) lived in houses owned by the 15

16 government. A large majority (92.34%) of the population lived in houses with electricity. Of these, 61.71% of these households received electricity from private providers, and 30.64% of the houses received electricity from the government. While a large number of houses appeared to be developed because of the presence of electricity, the story was not as bright with respect to sanitation. Only 26% of the households owned a latrine. 0.56% used a common latrine, while 73.35% of the houses had no access to a latrine. The network data is available at both the household and the individual level in adjacency matrices a nxn matrix, where n is the number of households in a village. These households are asked survey questions which help study network formation. Specifically, these questions to assess willingness to borrow are i) If you needed to borrow kerosene or rice, to whom would you go? ii) If you suddenly needed to borrow Rs for a day, who would you ask? And, the questions to assess willingness to lend are iii) iv) Who would come to you if he/she needed to borrow rice/kerosene? and Who do you trust enough, that if he/she needed to borrow Rs 50 for a day, you would lend it to her? (Rs 50 is approximately equal to a dollar.) 7 For each question and for each village, there are two adjacency matrices containing responses at the household level and the individual level. Rows and columns of the matrix represent an individual (or a household) in a village. Therefore each cell is the relationship between two individuals (or households). If a village has 396 households, then the adjacency 6 Fifty rupees is approximately equal to a dollar 7 The entire survey contains a lot more information, including information on from whom an individual would receive or to whom an individual would give medical advice, with whom a person would watch television, go to the temple, consult with for advice on personal matters, with whom they met in their free time, and who had helped them find a job. There is also information on whether an individual was a part of a savings group. 16

17 matrix for that village will have 396 x 396 cells. The cells take on a value of either 0 or 1. 1 indicates that the household in the row said that they would borrow from (or lend to) the household in the column. Data Management - Creation of variables A number of variables were created to summarize household borrowing and lending patterns, to measure village heterogeneity, and to indicate the prominence of the groups with which a household is identified. Dependent Variable - Borrowing and lending measure First, a measure to indicate the borrowing and lending patterns at the village level was created. As explained earlier, each cell in an excel file (matrix) represented one household. To generate a measure of the number of households from which one household was likely to borrow, a sum across each row of the matrix was obtained, as shown in Figure 1. In order to get a relative measure, the sum was divided by the total number of households in each village. A measure of the extent of the willingness to lend was generated in the same way. These measures are used in the empirical analysis to study determinants of the extent of borrowing and lending networks (for a summary of descriptive statistics on these measures, see Table 1). The borrow-pairs variable for this observation will be 3, a sum of the cells across the second row. Figure1: Example of an adjacency matrix. 17

18 Figures 2 and 3 illustrate the distribution of the borrowing and lending measures created for money and for rice and kerosene across the villages in our dataset. Even though our study employs percentage values for borrowing and lending, it might be useful to make note the borrowing and lending patterns in absolute terms. For lending and borrowing preferences in money and in rice and kerosene the average number of households that one household said they would borrow from or lend to were all in the range of The maximum number of households that a person said they could borrow and lend rice 8 from was lower than the number of households a person said they would borrow and lend money to. The highest average number of households among the four categories was found in the networks for borrowing money. The higher absolute number of households was reported for borrowing and lending money than for borrowing and lending rice and kerosene, which was interesting. In terms of percentage value, we see that on average, households networks include 1.58% % of other households in their communities (Table 1). 8 Henceforth, rice refers to rice and kerosene. In certain instances through this paper, rice and kerosene is mentioned as rice for the sake of brevity. 18

19 Figure 2: Lending pairs a percentage of households in the village one household said they would lend money to, across all the villages in the dataset. Figure 3: Borrowing pairs a percentage of households in the village one household said they would borrow money from, across all the villages in the dataset. 19

20 Figure 4: Borrowing pairs a percentage of households in the village one household said they would borrow rice and kerosene from, across all the villages in the dataset. Figure 5: Lending pairs a percentage of households in the village one household said they would lend rice and kerosene to, across all the villages in the dataset. 20

21 Independent Variables Indicators of Ethnic Identity and Heterogeneity The first indicator for ethnic diversity was the measure for ethnic fractionalization. For this, the Labar (2010) formula was used, as shown below ELF = 1 -!!!!!!!! where n i is the number of people in a particular ethnic group (religious or caste-based), N is the total number of people in the village and I is the number of ethnic groups in the village. The ELF takes on a value between 0 and 1. A measure of 0 indicates that there was no fractionalization and all people belonged to the same ethnic group. The closer it is to one, the more fragmented the community is. Second, the religious and caste-based polarization was calculated for each village using the specifications detailed by Montalvo and Reynal-Querol (2003).!!!!!.! )! π j POL = 1 - (!.!!!! Where π j is equal to n i /N. Table 1 displays the descriptive statistics for both indices in our data. The borrowing and lending behavior could also be driven by the ethnic identity of the person within the context of the community he belongs to. That is, a person might be borrowing or lending from another person based on the religious/caste group of the other person as opposed to, or along with, the total community level composition of ethnic groups. For this reason, a third 21

22 pair of measures is added the samereligratio and the samecasteratio that aim to capture the impulse to lend and borrow from a person of the same ethnic group. This measure is simply the number of people in the person s ethnic group divided by the total number of people in the village. Theoretical framework Consider a community populated by n ethnic groups, where the number of people in each group is greater than zero. Each household makes the decision of whether to borrow from another household in the village, or whether to lend to another household in the village. The community is binding, in that households can only borrow from and lend to people within the village. We assume a certain degree of implied reciprocity here. A person decides to lend to another person with the hope that at a future period, (time t+1), the other person will lend to them as well. When borrowing, the opposite is true the implication is that the borrower will lend to another person out if that person is in need of help in the future, because of the unspoken debt they have to the network now 9. A community with n=1 has only one group. The power threat hypothesis dictates that people in this community don t feel any threat to their position. People in this community feel most assured about their help being reciprocated, and most comfortable lending to others. Consequently, those borrowing from others also receive help easily. If the number of groups in the village increases, then the existence of another group in the village creates a power threat to 9 There may also be a sense of communal support independent of implied reciprocity, where individuals make lending decisions outside of consideration of any future economic and social benefit. What is the utility of such decision-making? It is harder to quantify or empirically observe, and is an area that future research can explore. 22

23 individuals of the first, and larger group. Individuals from the second group will also feel this threat because they are not in the majority and at a disadvantage. Individuals from both groups in this community have less faith or confidence that in a future event at time t+1, or t+2, their action would be reciprocated. As the number of people in the second group increases, the relative strength of the dominant group decreases, the distrust increases, and the strength of the informal insurance network will decrease. As the number of groups increase, and the relative sizes of one group with respect to another(s) changes (as the population becomes more heterogeneous), the contact among groups increases, building a different sense of community in line with the predictions of contact theory. That is, as the number of people from different groups increases, the probability of these individuals interacting via associational networks increases. Through this interaction, these groups lose the distrust they had in the event of reciprocity occurring in time t+1 or t+2. This trust could develop from demonstration of positive contract enforcement in the interactions among individuals from different groups in the village. We reach a point where the contact is sufficient to overthrow the negative effects of the power threat. At a certain n, there would have been enough exposure among the individuals such that an increase in the heterogeneity will no longer negatively affect the decision to lend to others in the community. Since n is infinite, that is, a community cannot infinitely keep increasing the number of ethnic groups, n could act as a constraint. For example, if the threshold at which contact theory takes over power threat in a certain community is 4, and the community has only 3 groups, then another factor has to play in for the effect of contact theory to be evident. This factor is the distribution of n. Therefore, we modify our previous statement to say that at a certain size and distribution of n, the community reaches the point where the effect of the power threat hypothesis is dominated by the effect of the contact hypothesis. 23

24 Together, our two variables for measuring ethnic heterogeneity closely capture the size, number and relational power of groups in a community. We should consider the behavior of the fractionalization and polarization indices next. The measure for ethnic diversity Fractionalization and its nature The ELF or ethnic fractionalization index takes a value between 0 and 1, based on the composition of the ethnic groups. It takes into account both the size and the number of groups. If a village has only one group, for instance, everyone belongs to the same ethnic group and the value of the ELF is 0. With two ethnic groups, 0.5 is the highest value that the variable can take and occurs when the groups are evenly split. Figure 6 depicts the range of ELF with the increase in the number of ethnic groups, assuming they are of equal size (which is when the ELF takes on the largest possible value). Number of Highest Possible value of groups ELF Figure 6: Upper range of values of the fractionalization index with different numbers of groups in a community If groups are of unequal size, the ELF will assume a value between zero and the highest possible value for that number of groups. For instance, if there were three groups, and the first one had 86% of the population, the second one had 11% of the population, and the third group had 3% of the population, then the ELF is , a value that is in between 0 and In this 24

25 example, there is a probability that two persons chosen at random in that community are from two different ethnic groups. The value moves closer to , the highest possible value for 3 groups, when the two non-dominant groups increase in size. It is important to note that in our data, the religion and caste ELFs take on different range of values. The largest value that the religion ELF can take is , because there are three different religions represented in the communities (Hinduism, Islam and Christianity). In actuality, the figures are much lower because of Hindu dominance and because many villages have only two religious groups, implying a maximum value of 0.5. The highest value of this variable in the data is With caste, there are five different groups General, Minority, OBC, Schedule Caste, and Schedule Tribe. Therefore, the caste ELF can range up to 0.8, and reaches a maximum value of in our data. Alternative measure for ethnic diversity - The polarization index The polarization index also ranges from zero to one. It attains its maximum at the bipolar symmetric distribution of groups in a community (Reynal-Querol, 2002). By going beyond a simple probability, that the fragmentation method measures, the polarization relates the possibility of conflict arising from relational dynamics between the groups. The polarization index is zero when there is only one group. With two groups, it gets closer to one as the distribution becomes more equal. For instance, a village in our dataset is divided 56.8% Hindus and 43.2% Muslims this is close to a divide, and the polarization in this village is The concept is easy to envision with two different groups. However, as the number of the groups increases, this becomes a little bit more complicated to picture. Three different examples of the distribution of caste groups within our data and the polarization values 25

26 they produce are provided in Figure 7 for better understanding of the behavior of the polarization index. Number of groups Polarization Distribution of groups Index A B A B A B Figure 7: Polarization values for different distribution of ethnic groups In each pair, the community marked A is closer to a more unequal distribution between the groups. Either one group has a clear majority or is significantly larger than the second largest group in the community. In these cases, the polarization is larger. The groups marked B have a more equal distribution and therefore lower polarization. It is interesting to note how the power threat component comes into play. With 5 groups, community B has one group with a clear majority. Community A has two pairs of groups that are closer to a bimodal distribution, and hence feel more power threat than communities in 5-B do. Therefore A has a higher polarization index. Polarization takes into account distribution how communities position themselves with respect to other groups in a multi-ethnic setting. In our data, there is a high correlation between religious polarization and fragmentation ( ), as well as caste polarization and fragmentation ( ). Because most (all but 7) villages in the data have either one or two religious groups, the fragmentation and polarization indices for religion essentially behave the same way and therefore have a very high correlation. There are a greater number of caste groups in our data; hence we see a relatively lower 26

27 correlation between the two variables there. It is interesting to look at the movement between the two indices in Figures 8 and 9. Two Indices: Religious Fragmentation vs. Religious Polarization religious polariza8on religious fragmenta8on Figure 8: Scatter plot of religious fragmentation index against religious polarization index 27

28 Two Indices: Caste Fragmentation vs. Caste Polarization caste polariza8on caste fragmenta8on Figure 9: Scatter plot of caste fragmentation index against caste polarization index The caste measures display an inverse-u shaped relationship, whereas the religion measures display a linear relationship. In both cases, polarization is at a maximum when fragmentation is half. Empirical Model 10 In order to determine the influence of ethnic heterogeneity on the extent of borrowing and lending networks, an OLS 11 regression that estimates the effect of the two indices on the borrowing and lending measures is used. 10 For a more detailed look at the choice of empirical model, please see Appendix B. 11 The model was also run using a GLS framework. However, results of the GLS framework were discarded because of a high Chi>2 specification on the Hausman test, which suggested that the GLS estimate would be unreliable and inconsistent. 28

29 The basic specifications can be found in equations 1-4 below. For each specification, the dependent variable, as mentioned earlier, is the fraction (like Chen (2007) and DeWeerdt s (2004) dependent variables) of households in a village from which one household said it would borrow (or to which it would lend). Our main variables of interest are the religion and caste ELF and polarization (relpol, castepol) measures. The squared terms for both ethnic heterogeneity measures are inserted in order to capture a non-linear relationship, like Chen (2007). The samereligratio and samecasteratio terms are identity variables. They indicate the percentage of the total population comprised by the household s ethnic group. These variables are present in order to capture the effect of behavior driven by trying to borrow or lend from a household in the same ethnic group as oneself. We expect that a person might lend to another person because of shared group identity. It is important to separately account for this effect in order to isolate the effect of the community level fractionalization. However, we expect the community level measures to have a stronger impact on the lending and borrowing measures. Montalvo and Querol (2003) acknowledge that the use of separate proportions for each religion cannot deal, for instance, with the effect on the (dependent variable) of diversity and latent conflict among religious groups inside a country. They advocate the use of the fractionalization index. The last three variables are indicators of wealth the nature of ownership of the person s house, whether they have electricity, and whether they have a latrine. These were all recorded as categorical variables from the survey, and were introduced into the model as dummy variables to control for any potential wealth effects on network formation. 1. lendriceratio = β 1 relelf(relpol) + β 2 casteelf(castepol) + β 3 relelfsq(relpolsq) + β 4 casteelfsq(castepolsq) + β 5 samereligratio + β 6 samecasteratio β 7 house + β 8 elec + β 9 latrine + ε 29

30 2. borrowriceratio = β 1 relelf(relpol) + β 2 casteelf(castepol) + β 3 relelfsq(relpolsq) + β 4 casteelfsq(castepolsq) + β 5 samereligratio + β 6 samecasteratio β 7 house + β 8 elec + β 9 latrine + ε 3. lendmoney= β 1 relelf(relpol) + β 2 casteelf(castepol) + β 3 relelfsq(relpolsq) + β 4 casteelfsq(castepolsq) + β 5 samereligratio + β 6 samecasteratio β 7 house + β 8 elec + β 9 latrine + ε 4. borrowmoney= β 1 relelf(relpol) + β 2 casteelf(castepol) + β 3 relelfsq(relpolsq) + β 4 casteelfsq(castepolsq) + β 5 samereligratio + β 6 samecasteratio β 7 house + β 8 elec + β 9 latrine + ε The information on caste was unavailable for the first 25 villages in the household data. However, the individual dataset had caste information for all the villages in the dataset. Using the individual level data, a proxy for the missing village fractionalization values was calculated. The obvious drawback of using these values is that it is from a non-stratified sample of the population. However, we run the regression using this line of analysis because it is the closest estimate we have of the actual measure of caste ELF from the data, and we find that the correlation between the caste ELF calculated at the HH level and the proxy variable for the remaining 50 villages is Results Results for the regressions are listed in tables 3 (borrowing rice and kerosene), 4 (lending rice and kerosene), 5 (borrowing money), and 6 (lending money). The regressions were run in five different sets i) with religious fractionalization of all 75 villages in the mix, ii) with caste and religious fractionalization of 50 villages, excluding those villages that did not have perfect 30

31 data on the caste of the households, iii) for 75 villages using the proxy caste ELF measure constructed from individual level data, iv) with religious polarization of all 75 villages in the mix, v) with caste and religious polarization of 50 villages. In the first specification using the ELF (fragmentation measures), we observe significant results for the impact of religious heterogeneity on the formation of informal insurance networks. As the community moves from being homogenous to heterogeneous, people are less likely to want to lend money or rice and kerosene to others in the community. They also display a lesser tendency to borrow from other people as the heterogeneity increases. Because the quadratic term for religion ELF is also significant, this relationship is non-linear. The relationship therefore switches from being inverse to direct after a certain threshold level. The points at which this switch occurs is in the range of for our data (Table 7). Among the four different categories of borrowing and lending, the lowest turning point (0.239) was observed for lending money in model (2). The highest turning point (0.274) was observed for borrowing money in model (3). In our data, the range of the religion ELF across villages is Therefore, when the community turns from completely homogenous to heterogeneous we observe a turning point in the relationship. The marginal change in the percentage of households a given household would lend to or borrow from reduces with an increase in the religion ELF from 0 to the turning points for each of the different models. Graph 1 helps picture this relationship for one of the models. The direction of the relationship is the same for all of the others. 31