Why Does Ethnic Diversity Affect Public Good Provision? An Empirical Analysis of Water Provision in Africa

Transcription

1 Why Does Ethnic Diversity Affect Public Good Provision? An Empirical Analysis of Water Provision in Africa Ken Jackson October 5th, 2007 Abstract The lack of provision of basic services such as piped drinking water is a major barrier to economic development in many countries around the world. However there is little consensus about the mechanisms that limit the ability of communities to provide these goods. In this paper I investigate the role of ethnic diversity as a potential barrier to the provision of public goods. The existing literature focuses on two primary mechanisms through which local ethnic diversity reduces the supply of publicly provided goods - either ethnic groups have diverse preferences or the local institutions that manage inter-ethnic relations are weak. The similarity in the aggregate predictions of the two theories stands in stark contrast to the differences between their policy implications thus providing the motivation to differentiate their effects. In this paper I embed both mechanisms within a single theoretical model and derive empirical implications. I then test the implications of the model for the effects of ethnic diversity on piped water provision in Sub-Saharan Africa. With a series of household surveys covering 15 countries I am able to test theoretical predictions at both aggregate and household levels. The results indicate that ethnic diversity has signficant negative effects on the provision of piped water and ineffective local institutions are the primary channel for this effect. I would like to thank participants in the UBC development and empirical lunch seminars for their helpful advice. In particular Chris Bidner, Chris Barrington-Leigh, Mukesh Eswaran, Patrick Francois and Nathan Nunn have provided numerous valuable comments and ongoing support. This paper is work in progress so any comments would be very much appreciated. - kjackson@interchange.ubc.ca, Website - grad.econ.ubc.ca/kjackson 1

2 1 Introduction The international development community should speak of the Big Five development interventions that would spell the difference between hunger, disease and death and health and economic development. (Sachs 2005, p.235) The Big Five interventions discussed by Sachs are: 1. Agricultural improvements (related to poor credit markets) 2. Investments in basic health 3. Investments in education 4. Power, transport and communications services 5. Safe drinking water and sanitation There are two common factors that are present in each of the elements on this list. First, each of these goods requires large upfront expenditures. And second, the list describes essentially all productive government activity in most countries of the world 1 A related question is therefore why aren t the governments of developing countries able to provide the basic services necessary for development? One factor that has been suggested as potentially limiting the effectiveness of government is social heterogeneity, with special attention focused on ethnic diversity over the past 10 years. In particular, ethnic diversity is critically important within Sub-Saharan Africa where the greatest levels of diversity are found and the problems of underdevelopment are particularly intractable. Further complicating the issue is that there are two primary mechanisms that potentially link ethnic diversity and poor public good provision and while their aggregate predictions are similar their policy implications are not. The first mechanism, initially suggested in Alesina, Baqir and Easterly (1999)(henceforth ABE) 2, hinges on variation in preferences between different ethnic groups. The form of the resulting public good is therefore a compromise that does not fully satisfy any group. As a result the optimal level of provision is lower in diverse communities. To the extent that the preferences approach is correct, and taking settlement patterns as given, the resulting level of provision is efficient. As such there are few practical policy implications to be considered by either local officials or the development community. 1 The two main categories of government spending not on the list are military spending and transfers between groups - neither of which are usually considered productive. 2 Models based on this mechanism are also found in Alesina and La Ferrara(2000) and Kimeyi (2006). 2

3 The alternative approach, as formulated in Miguel and Gugerty (2005)(henceforth MG), is based on the dea that the efficiency of coordination is higher within ethnic groups than between them. Diverse communities thus face higher costs of coordinating the provision of public goods and this is the link between ethnic diversity and poor governance outcomes. This approach suggests that reforming local institutions to reduce the cost of inter-ethnic coordination could significantly improve the provision of public goods and therefore development outcomes. The existing empirical literature has focused on measures of aggregate public good provision and thus are unable to differentiate between these approaches. It is therefore the goal of this paper to not only determine the aggregate effect of ethnic diversity on public good provision but to provide insight into the likely underlying mechanism. With this in mind the theoretical section of the paper embeds these two approaches within a single model that provides substantively different empirical predictions depending on whether the mechanism at work is primarily based on differences in preferences or poor institutions. The aggregate predictions of this model are very similar to those previously discovered and are effectively independent of the underlying mechanism. However, at the level of the individual household the model predicts different effects from the two approaches. We can therefore test the relative importance of each mechanism using household data on access to public goods. An additional factor that is vital to assessing the true impact of ethnic diversity on public good provision is the geographic scale at which diversity measurement is taken. The relevant geographic scale of diversity in each of the candidate theories is the distance at which investments in the public good provide substantial benefit to outlying areas. In the preferences approach if two neighbouring communities do not benefit from the investments of the other then there is no effect of diversity between the two towns. Similarly, in the institutions approach there would be no benefit to the two communities coordinating their activities and thus aggregate diversity would not increase governance costs. Diversity, as measured locally, would affect provision in each community and thus would have an effect on the average provision across the two communities. Alternatively, measurement at too narrow a scale would ignore potentially relevant diversity that has an effect on the underlying decisions. It is therefore critical that the scale of measurement for a diversity measure be related to the scale of provision associated with the public good. The empirical section of the paper uses household data from 15 countries of Sub-Saharan Africa to analyze the relevance of ethnic diversity on the provision of piped drinking water. The provision of piped water is an inherently local good due to high costs of water transmission. As discussed in Clark and Stevieciteyearclark81 the provision of water over ranges larger than approximately 10 miles faces sharply rising average costs. While the relevant measure in Africa could be slightly different due to 3

4 various factors 3, in this analysis I conduct the primary analysis at a local scale (10km) as suggested by theory and then demonstrate the effects of the specification bias associated with measuring diversity over larger or smaller areas. The aggregate results demonstrate that ethnic diversity has a significant negative impact on the provision of piped water in Sub-Saharan Africa. The fraction of households accessing piped water in homogeneous communities is increased by 30% from the most diverse communities. These results are therefore consistent with the existing empirical literature on ethnic diversity and public good provision. However the results are perfectly consistent with both theoretical approaches considered in the literature and thus do not resolve the central issue in this paper. To differentiate between the competing theories I consider the implications of each approach at the household level. If different ethnic groups vary in their preferred form of public goods then one would expect the public good that is chosen will align more closely with dominant groups in the community than smaller groups. Thus at the household level individuals in majority groups would be more likely to access the public good than those in minority groups. In contrast, if the public good is equally desirable to all groups then being a member of the majority group should have no significant effect. The results of the household analysis indicate that while households in more homogeneous communities are more likely to access piped water there is no additional advantage for those households that are members of the largest local ethnic group. The evidence thus supports the institutions story in explaining why ethnic diversity affects the provision of piped water in Africa. 4 These results therefore suggest that improving the local institutions that govern inter-ethnic relations would likely have beneficial impacts on the provision of public goods. The existing literature on ethnic diversity and the provision of public goods has developed quickly over the past decade. Starting with Mauro (1995) and Easterly and Levine (1997) a large literature developed analyzing the effect of ethnic diversity on national outcomes. It found significant negative effects of ethnic diversity on economic growth, with the primary channel being the effect on government policy 5. Easterly (2001) and Collier (2000,?) demonstrated that effective institutions eliminate any 3 For example, poor provision of electricity may create higher prices for the pumping of water and thus reduce the optimal scale of provision. Alternatively poor access to appropriate water sources may increase the value of transmitting water. 4 These results also excludes a potential third avenue through which ethnic diversity would affect public good provision. If public goods are controlled by a single group that limits access of other groups then one would find the aggregate prediction that diversity reduced public good provision. However, as long as the likelihood that a group is dominant is correlated with their relative size this would have similar household consequences to the preferences theory. 5 Alesina et al (2003), Fearon (2003) and Posner (2004) refined the measurement of ethnic diversity - collectively demonstrating that the result is robust to variation in the way ethnicity is measured 4

5 negative effects associated with ethnic diversity. Focusing directly on the provision of public goods La Porta et al. (1999) provided evidence that ethnic diversity negatively affects literacy and public health, while having less robust effects on educational attainment and infrastructure quality. Kuijs (2000) analyses the impact that ethnic diversity has on both the quantity and quality of public expenditures on education and health. Ethnic diversity is found to negatively affect the quality of spending in both cases and the quantity of spending in the case of public health. 6 At the local level, Alesina et al.(1999) provided the first analysis of the effects of local diversity on the provision of public goods. The empirical results considered U.S. city and county expenditures on a variety of public goods and found significant variation in the effects of ethnic diversity on different goods. As an example, ethnic diversity was found to lower expenditures on roads and healthcare, increase expenditures on policing and have no effect on fire services. In western Kenya, Miguel and Gugerty (2005) find significant negative effects of ethnic diversity on the provision of education and well water. In further research from developing countries, Bardhan and Dayton-Johnson (2002) summarizes the results of a series of papers in the area of diversity and agriculture projects and finds that most (but not all) report negative effects of ethnic diversity on local cooperation and Khawaja (2006) also finds significant negative effects of ethnic diversity on maintenance of public projects in Pakistan. In each of these papers the geographic scale associated with the public good is essentially the same as the scale of diversity measurement. In most instances this might be expected, as a literature in public finance has demonstrated that there are potentially significant gains to having the provision of public goods organized by the smallest government unit that is large enough to account for the majority of the benefits associated with the public good(oates 1972, Besley, Pande, Rahman, and Rao 2004). Therefore in the optimal case it is sufficient that measurement occurs at the scale associated with the governing body supplying the good. However in Banerjee et al.(2005) and Banerjee and Somanathan (2007) measurement occurs at the federal district level in order to allow for the use of historical diversity measures. These districts each contain multiple local communities and it is apparent that investments in one town do not necessarily benefit those living elsewhere in the district. 7 6 In the model developed below this outcome would be expected if inter-ethnic spillovers are relatively low in the case of education and relatively high in the case of health measures. I discuss this further at the end of section 2. 7 The measure of provision is the fraction of towns that access the public good - implying that some towns in a district do have access while others do not. 5

6 In the first paper the authors find significant negative effects of social diversity 8 while in the second the effects remain significant but small. In light of the measurement issue mentioned above and the empirical results provided below it seems possible that this variation in results is due to the specification error caused by data limitations. Caselli and Coleman (2002) provide a theoretical justification for focusing on forms of social heterogeneity that are relatively expensive to change such as ethnicity. However the model considered here does not exclude the possibility that groups may form around alternative social constructs such as religion. In addition to the results related to ethnic diversity above, Alesina et al. (2003) further estimates the impact of religious diversity on economic growth and finds no significant impact. In the empirical section I also consider the possible role for religious diversity and similarly find no evidence that indicates heterogeneity along this dimension is important to the provision of public goods in Sub-Saharan Africa. 9 Section 2 develops the theoretical model and further discusses the predictions of the model. Section 3 presents the empirical specification with which the theoretical implications are tested while section 4 contains a description of the data used in the analysis. Section 5 contains the primary results. Section 6 contains conclusions and directions for future research. Proofs and robustness checks are confined to the appendix. 2 Model The existing literature on ethnic diversity has considered two primary mechanisms for how ethnic diversity could affect the provision of public goods. First, different ethnic groups could want different public goods. In the model of Alesina et al. (1999) (ABE) this is the relevant source of conflict associated with diversity. The provision of public goods in diverse communities involves lower average satisfaction for any level of funding and therefore the optimal level of funding is lower. Alternatively, Miguel and Gugerty (2005) (MG) develop a model in which preferences do not diverge and the lack of effective inter-ethnic institutions induces a free-riding problem between ethnic groups. Ethnic diversity is thus associated with lower provision of public goods because of the differential effectiveness of intra-ethnic institutions as opposed to inter-ethnic ones. The model developed here embeds the 8 The paper combines measures of caste and religion in determining diversity and thus it is not clear that the results are directly comparable to papers only considering ethnic differences. 9 As the empirical section only considers countries in Sub-Saharan Africa it is not possible to conclude anything about the potential importance of religion in other areas of the world - or in countries of Africa not contained in the study. 6

7 primary features of both models and therefore allows for different empirical predictions for different situations based on the nature of the public good and the institutional environment. 2.1 Environment A community provides a single public good funded by the contributions of local households. 10 The community has a population of mass N, divided among a total of E ethnic groups with p e defining the population share of ethnic group e. Households receive utility from private consumption and the public good: u i = W x i + µ ln X i (2.1) where x i is the household s contribution to the public good and X i is the benefit the individual receives from the public good. As a simplification 11, it is assumed that ethnic groups are able to fully overcome the internal free-riding problem and thus all individuals will make contributions to maximize group utility. 12 Individuals from the same ethnic group face identical decisions and therefore I will focus on solutions that are symmetric within groups and we can therefore denote x i as x e. The benefit received from the public good depends on the contributions of all other individuals but the contributions from one s own ethnic group may potentially have a greater effect. X e = p e Nx e + α p e Nx e (2.2) e e The parameter α [0, 1] captures the extent to which the preferences for a public good vary across ethnicities with α = 1 indicating no variation in preferences. Each group is assumed to choose their contributions taking the actions of all other groups as given. However the institutional environment may provide incentives to consider the utility of other ethnic groups. Each group takes the contributions of other groups as given and chooses x e 0 to solve: E U e = max (1 τ)p e N(W x e + µ ln X e ) + τ p j N(W x j + µ ln X j ) (2.3) x e 10 Alternatively funding from higher level governments could be available but responsive to local lobbying efforts. For the results below it would be sufficient that both the size and form of the public good provision are sensitive to local activities. 11 This is consistent with the model in MG and not relevant to the model of ABE in which institutions are essentially irrelevant. j=1 Alternatively, it could be assumed that the institution governing intra-ethnic coordination is less than perfect. To the extent that the ability of groups to overcome the within-group public good problem is not affected by group size the solutions provided here would not be changed in any significant way. 12 This could be due to an intra-ethnic sanctioning technology as in MG or due to intra-ethnic altruism as in (Vigdor 2004). 7

8 where τ [0, 1]] parameterizes the effectiveness of inter-ethnic institutions. With τ = 1 all group decisions will be the same as those that would be made by a hypothetical benevolent social planner. 2.2 Extreme Cases Before proceeding to the main propositions I present solutions to the extreme cases as these effectively demonstrate the different effects of the two primary mechanisms Divergent Preferences In the event of completely divergent preferences (α = 0) there is no effect of the institution, τ. Each group invests on their own behalf and each group invests until: X e = µp e N (2.4) The average benefit is therefore: X = µ(1 F )N (2.5) where F is local ethnic fractionalization. This result is efficient regardless of institutional quality and in this case there is no practical policy implication in response to the inefficiency caused by ethnic diversity. The reason for this is that contributions by one group have no impact on the welfare of any other group and hence there is no reason to coordinate Complete Spillovers In the event of complete spillovers (α = 1) the results depend critically on the quality of inter-ethnic institutions. If inter-ethnic institutions are non-existent (τ = 0) only the largest group will contribute and all other groups will free-ride. The benefits to all groups are the same and are determined by the size of the largest group. X e = X = µp 1 N (2.6) However if inter-ethnic institutions are complete τ = 1 then all individuals contribute 13 and: X e = X = µn (2.7) As will be confirmed below the only case in which diversity does not have an effect on public good provision is when neither the preferences or institutions approach is valid. 13 In the case of τα = 1 the model faces multiple equilibria as it is irrelevant which group(s) contribute. The predictions of the model do not depend on which individuals contribute and therefore this mulitiplicity of equilibria is irrelevant to this analysis. 8

9 2.3 General Case In a general solution we must consider the non-negativity constraint on public good investment by each group. However, lemma 1 significantly restricts the set of possible cases that require analysis. Lemma 1 A member of a larger group will never contribute less than a member of a smaller group. They will contribute strictly more unless α = 0, τα = 1, or they do not contribute at all. In addition, the largest group will always contribute. Proof - see appendix. The effect of this lemma is that one can order groups by size and restrict analysis to cases where groups 1 through K contribute and groups K+1 through E do not. In all following results I assume that p i p i+1, i < E 1. Following directly from the proof of lemma 1 is proposition 1. Proposition 1 For α < 1,if p 1 > p j, X 1 > X j. For α = 1, X 1 = X j. Proof - see appendix. If there is a single largest group they always contribute more than any other group and therefore the public good is directed more closely towards their preferences. This proposition provides the primary difference between the preferences approach and the institutions approach. In the preferences model α < 1 and individuals in the largest group should receive a larger benefit from public goods than other individuals. Alternatively, if all individuals benefit equally from the public good then any effect of ethnic diversity on public good provision must be due to poor institutions. For our next result I define average public good provision as the weighted average 14 of group benefits: E X = p e X e (2.8) e=1 Proposition 2 If τα < 1 increasing the population share of the largest group (at the expense of any ( ) X other group) will strictly increase average public good provision p 1 > 0. Proof - see appendix. In general there are two effects associated with a reduction in diversity. First, the individual(s) that move to the largest group receive a large direct benefit associated with improved access to public goods. Second, if the individual(s) were not previously contributing their contributions 14 The proposition is unchanged if we define X = E e=1 pe ln Xe. 9

10 increase the total level of spending on the public good. The proof demonstrates that the result of these effects in unambiguously positive on the average level of public good provision. In general increasing the size of any group that was contributing at the expense of any smaller group will increase average public good provision. However, whether the second group is contributing depends on α. If α < p2 p 1 then the second largest group will also contribute. This leads to the following corollary: Corollary 1 If α < p2 p 1 then increasing the population share of the second largest group (at the expense of a smaller group) will strictly increase average public good provision. Proof - the proof of proposition 2 does not depend on the identity of the group increasing in size. It is sufficient that the increasing group is contributing to the public good. This is the case when α < p2 p 1. Thus if variation in preferences is high enough, then holding the size of the largest group constant and increasing the size of the second largest group should increase aggregate provision. Further, as α is reduced the number of groups that will contribute increase. For α = 1 the relevant measure of diversity (or homogeneity) is the size of the largest group. When α = 0 the relevant measure is fractionalization. contributing groups. In between these extremes the optimal measure is the fractionalization between 2.4 Quantity vs. Quality of Spending Kuijs (2000) indicates the potential for variation in the effects of diversity on the quality of spending versus the quantity of spending. This model provides one channel to explain why these effects would be different in specific cases. The quantity of spending increases as diversity increases the number of individuals contributing. When α is low many small groups will be contributing and therefore changes in diversity will not increase total spending. The effectiveness of spending is determined solely by α. When α is high spending benefits all individuals whereas when α is low more spending is required to benefit each group. Diversity will therefore have a larger effect on the quantity of spending when variation in preferences are low and a larger effect on the efficiency of spending when there is great variation in preferences. In Kuijs (2000) spending on education is not affected by diversity but the efficiency of spending is significantly affected. This is therefore consistent with education being a good where ethnic groups have large variation in preferences. Alternatively, both spending and the efficiency of spending on public health programs were reduced by diversity - indicating that health outcomes may face lower variation in preferences. 10

11 3 Empirical Specification The empirical section of this paper tests the implications of the model for the provision of piped water in Sub-Saharan Africa. The existing empirical literature on ethnic diversity and public good provision has measured the impact of local diversity (usually measured as fractionalization) on the aggregate provision of public goods at the community level. The first results reported are similarly conducted at the aggregate level to provide estimates that are directly comparable to the existing literature. The model developed above indicates that if either preferences vary across ethnic groups, or local institutions are not powerful enough to overcome inter-group free-riding, or both, then the aggregate provision of public goods will be lower in communities that are more diverse. The ideal specification is therefore: X t = β 0 + β 1 D t + γ 1 G t + γ 2 E t + ɛ t (3.1) where D t is the optimal measure of diversity as discussed above, G t, E t are vectors of geographic and economic controls, respectively, that capture the relevant cost-benefit factors that may affect the decision to invest in a piped water system. Two problems arise from this description. First, in the available data there is no direct measure of public good provision. To resolve this I calculate the fraction of households that access the public good, and thus the specification will be correct if this fraction is an increasing function of the aggregate level of the public good. Second, the model does not provide a single, optimal definition of diversity. For our purposes both the size of the largest group and fractionalization would be potential options and I provide results using each measure 15. At the aggregate level finding β 1 < 0 would indicate that ethnic diversity is having an effect, though the channel for that effect would be uncertain. In order to differentiate between the two mechanisms I then test proposition 1 using a householdlevel specification. X i = β 0 + β 1 D t + β 2 M i + γ 1 G t + γ 2 E t + γ 3 H i + ɛ i (3.2) where in addition to the previous regressors, M i is a measure for the household that takes the value 1 if individuals in the household are the same ethnicity as the largest local group, and H i is a vector of household controls. If the variation in preferences between groups has an impact on the provision of piped water then individuals in the largest group in a community should have better access to the public good and therefore in addition to β 1 < 0 one should expect β 2 > 0. Only if α = 1 does the model suggest that β 2 = As noted below, these measures are highly correlated. 11

12 3.1 Accounting for Scale The analysis above effectively defines a public good as one for which investments by one individual benefit another individual. An additional factor that varies across public goods is the geographic distance over which other individuals will benefit from a single investment in a public good. For an ethnic group maximizing group utility, investments are only affected by the number of individuals in the ethnic group that will benefit from the investment. It is therefore critical to measure diversity at the scale of geographic spillovers associated with the public good. The importance of the scale of measurement is caused by the problem of aggregating diversity measures at larger districts. Thus the average wealth of a province is the weighted average of the average wealth of each community within the province. However the relationship between aggregate diversity and local diversity is far more complicated. It is therefore necessary to match the diversity measure to the intended mechanism. Within the context of the model above, the mechanism for ethnic diversity is limited to the region over which an individual s investments would benefit other members of their ethnic group. For a good such as piped water, these benefits are severely restricted outside a relatively short distance. The reason for this is that pumping water over long distances is extremely inefficient and therefore water projects face sharply increasing costs outside of a few kilometres. 16 The results in table 6 below demonstrates the potentially severe specification error that can occur when measurement occurs at an inappropriate scale. 4 Data The primary data source for this analysis is the Demographic and Health Survey conducted by the Measure DHS project in 75 countries. This analysis is conducted on all countries of Sub-Saharan Africa for which sufficient data is available - resulting in 15 countries. 17 For each country I use the last available survey that contains sufficient data, resulting in survey years that vary from 1994 in Cote D Ivoire to 2004 in Cameroon. The complete dataset contains the results for approximately 180,000 individuals in 100,000 households across 5,700 survey clusters. As mentioned above it is critical to measure diversity at a level appropriate to the public good under analysis. In the case of water the appropriate scale is approximately 10km and therefore I 16 Clark and Stevie (1981) find that the efficient scale for water projects in the United States was approximately 9 miles. In Africa where electricity provision is much poorer than the U.S. the cost of pumping is almost undoubtedly higher and thus efficient scales would be expected to be even smaller. 17 Excluded countries are those for which either ethnicity data or geographic data are not available. 12

13 Countries in Dataset Mali Niger Senegal Chad Burkina Faso Benin Togo Cote d'ivoire Ghana Cameroon Central African Republic Ethiopia Kenya Malawi Namibia create arbitrary districts that are 10km x 10km across the entire sample and combine the data from all households within each district. This level of aggregation combines the 5,700 survey clusters into 3,300 districts with individual districts containing between 1 and 67 clusters. Additionally, this is partly done out of necessity as local government borders are not consistently available across all countries. Further, even if these boundaries were available it is not clear that local governments would not take into account the benefits of their investment that fall outside their borders. This is in part due to the dual nature of African politics with significant influence remaining for the traditional leadership structure in many countries. For example, in Ghana, Owusu-Sarpong (2003) points out that there are two sets of local political entities that affect development projects. the central government of Ghana,...,can safely carry out its development projects only by relying on a strong cooperation between the two complementary local political entities: 13

14 the institutionalized local government structure and the perennial traditional authority structure; for the latter remains close to the heart of the people. (Owusu-Sarpong 2003, p.34) 4.1 Dependent Variables For the district level specification, the primary dependent variable is the fraction of households that report receiving their drinking water from a piped source. This includes both households with water piped into their house or compound along with those that access water at public taps. The alternatives include well water, boreholes, streams or rivers and rainwater. The necessary assumption is simply that the provision of a piped water system generally requires a more significant infrastructure cost than each of the alternatives and is thus more likely to be affected by the ability of communities to organize funding activities. For the household specification, the dependent variable is a binary variable indicating whether the household accesses water from a piped source. 4.2 Ethnicity The diversity measures contained in this study rely on the ethnicity data reported in the measure DHS surveys in each country. In the survey each respondent is asked to select their ethnicity from a list of potential ethnicities with Other always as an included option. As shown in table 1 the number of ethnicities listed varies across countries from a low of 7 in Togo to a high of 66 in the Ivory Coast. In contrast to the existing literature on ethnic identity and local public goods, this paper combines data from countries with widely varying ethnic structures. An example of this is the contrast between Burkina Faso where the Mossi are the dominant group in size and wealth as opposed to Niger where the Haoussa are the largest group but are relatively poor. Choosing an ethnicity from a list may have a tendency to force individuals to make a choice that does not entirely fit their identity. To the extent that this biases the diversity results, and assuming that this error is unrelated to the provision of public goods, the measured effect of diversity will tend to be biased towards zero. This problem may have less significance in Africa as opposed to other regions due to the historical prevalence of ethnic identity as a tool of colonial rule. One form of evidence for this is the strong correlation between the results of the household ethnicity survey and recent alternative data sources of ethnic diversity in Africa. Not only do national measures of diversity created using this dataset agree closely with published data in other sources, such as Alesina et al. (2003) and Fearon (2003), table 2 uses Kenya as an example and provides striking evidence 14

15 Country # Ethnicities Largest Ethnicity Frac.Large Group Wealth Benin 10 Fon 45.5 Average Burkina Faso 11 Mossi 53.6 Above Avg. Cameroon 50 Bamileke 19.8 Above Avg. CAR 9 Gbaya 28.5 Below Avg. Chad 14 Sara 26.4 Average Cote D Ivoire 66??? 15.6 Above Avg. Ethiopia 56 Amharra 33.2 Above Avg. Ghana 9 Akan 52.8 Above Avg. Kenya 15 Kikuyu 23.3 Above Avg. Malawi 10 Chewa 32.3 Below Avg. Mali 10 Bambara 33.1 Below Avg. Namibia 10 Oshiwambo 46.4 Below Avg. Niger 10 Haoussa 60.6 Below Avg. Senegal 12 Wolof/Lebou 36.5 Above Avg. Togo 7 Adja/Ewe 47.2 Above Avg. Table 1: Ethnic Structure of the similarities between the individual components of the different measures 18. Where there is variation between the survey and published national level resources the survey tends to provide more options for survey respondents. This is important as local diversity may be significantly affected by the presence of groups that have a relatively small presence nationally. 4.3 A Measure of Diversity The measure of diversity used throughout the empirical literature on this topic is ethnic fractionalization. The model above predicts that the correct measure of diversity will depend on the public good with measures varying from fractionalization to the population share of the largest group. In practice the results are unaffected by the particular measure of diversity as the correlation between the size of the largest group and the level of ethnic fractionalization is Similar tables for other countries are available upon request. It seems likely that the consistency of how groups are defined across countries in Africa is the result of the colonial powers institutionalizing ethnicity to manage local areas. While there is evidence that this strategy was more commonly followed by the British it occurred to some extent in all colonial settings (Blanton, Mason, and Athow 2001). 15

16 Ethnic Group % of Nat.Pop.(DHS) Alesina et al. (2003) Fearon(2003) Kikuyu Luhya Luo Kamba Kalenjin Meru w/kikuyu Kisii Mijikenda/Swahili Somali Masai Embu 1.6 w/kikuyu Turkana Taita/Taveta 1.2 Kuria 0.6 Other Table 2: Ethnicity in Kenya An alternative measure of diversity that is more commonly seen in the literature is polarization. This measures the potential for inter-ethnic conflict and thus reaches a maximum with two equally sized groups. In the context of the model developed here it is clear that holding the size of the first group constant increasing the size of the second group should not have a negative effect. The results in table 5 are thus consistent with the model developed here. Table 3 lists summary statistics for piped water and diversity within Africa. As should be expected diversity is lower in rural areas while access to piped water is substantially lower. Diversity is naturally affected by migration and this creates a potential endogeneity bias. In the simplest story migrants may be drawn to areas because of high levels of public good provision and to the extent that in-migration tends to increase diversity this would create a positive relationship between diversity and the public good. This would generally bias the results towards zero. However to control for in-migration I instrument for all diversity measures using the same statistic as calculated only on those households that have lived in their current area for at least 20 years All individuals surveyed report the length of time they have lived in the area. I use all households for which the maximum tenure is either greater than 20 years or reported as Always. 16

17 Variable All Districts Rural Urban Fraction w/ Piped Water (0.346) (0.225) (0.298) Share of Largest Group (0.245) (0.198) (0.236) Share of 2nd Largest Group (0.117) (0.123) (0.090) Ethnic Fractionalization (0.285) (0.242) (0.245) Number of Districts 3,349 2, Notes: Standard deviations in parentheses. Table 3: Summary Statistics - District Statistics Alternatively it is possible that poor public good provision will lead to out-migration. Without historical data regarding ethnic diversity it is not possible to control for this factor as I do for inmigration. However it is also not clear what effect out-migration would have on local ethnic diversity. It is possible that individuals in minority groups will feel less tied to the community and would be more likely to emigrate. To the extent that this story is true the results will be biased towards zero and therefore should be taken as a lower bound. In general the average household reports living in their current area for over 30 years and thus to the extent that diversity exists it is relatively historical. The endogeneity of diversity is likely to be less of a problem in rural areas as opposed to urban areas and therefore I estimate effects both generally and over the rural portion of the sample. As approximately 70% of the sample is rural the differences between the two sets of estimations are small. 4.4 Geographic Controls The use of a piped water system may be strongly affected by geographic variables that we can control for in this study. Elevation, distance to a river, distance to the ocean and latitude 20 are included in the regressions. The distance to a river could have multiple effects as being close to a river would possibly reduce the cost of installing a water system but would also dramatically reduce the associated benefit. In addition to these I control for country and province fixed effects and thus further reduce 20 This is actually the absolute value of latitude as Namibia and Malawi are south of the equator. 17

18 the potential error associated with uncontrolled geographic variation. 4.5 Ethnic Controls African countries tend to be heavily centralized and funding for water projects may be strongly affected by political affiliation in the community. For this reason the presence of a large number of individuals of the dominant national ethnicity may affect the viability of development projects. To control for this effect I include the fraction of households in the community belonging to each of the three largest ethnic groups in the country. 21 While I have not assumed variation in demand for piped water across ethnicities in the household specification I do control for the ethnic identity of the household. 4.6 Other Controls Naturally, wealthier regions will be able to fund public projects such as piped water systems and thus controlling for wealth is necessary. The available wealth data is constructed from household asset lists and then standardized for each country to be mean zero with a standard deviation of 1. There are two possible sources of endogeneity created by this measure. First, the construction of the wealth measure may be affected by household access to piped water and thus public good provision may be associated with increases in the wealth measure (without an associated increase in actual wealth). Second, the existence of local public goods may increase local wealth. For example, quick access to clean piped water may reduce the burden on the household of collecting drinking water and thus allow individuals in the household to devote their time to more productive activities. To the extent that each of these is true they will induce a false, positive correlation between access to piped water and wealth. If the hypothesized relationship between diversity and the public good is true then this result will tend to bias the measured effect of ethnic diversity towards zero. To partially resolve the endogeneity problem I instrument for wealth using stock variables that go into the measurement of household wealth. The primary instrument that I use for this purpose is the flooring material in the household. This is unlikely to be directly affected by the source of local drinking water and thus avoids the measurement problem mentioned above. In addition, households do not tend to change the flooring material of their house frequently and therefore the stock nature of the variable will tend to reduce the immediacy of the endogeneity issue. As shown below in the results, instrumenting for wealth does not dramatically affect the results and this fact is robust to the choice of instruments These are country specific thus adding 45 variables to the regressions. 22 Though this choice is limited to alternative measures that will are subject to similar criticisms. 18

19 Population density is likely to strongly affect the efficiency of a piped water system. In this study I do not have a reliable measure of population density at a small enough scale for the local results highlighted here. Urban/rural status is used as a general predictor of density. In general density is associated with high levels of ethnic diversity and piped water efficiency and not controlling for density should therefore bias the results towards zero. In addition this problem is likely to be far less significant in rural areas that will not vary in density to the same degree as urban areas. A potential concern related to diversity is that diverse communities may be more likely to contain recent migrants or younger households and thus community organization may vary for reasons only indirectly associated with diversity. I therefore control for the length of residency reported by each household 23 and the average age of the household head. At the household level previous research (Thomas 1990) has indicated that the sex of the household head affects household decision making and therefore I control for this at both the household level and as a community average. 5 Results 5.1 Aggregate Specifications Primary Specification Proposition 2 indicates that more diverse communities should have lower aggregate levels of public good provision regardless of the mechanism through which ethnic diversity has an effect. While the model does not provide a single measure of diversity there are two possible candidates related to the two extremes of the model. These are fractionalization 24 and the population share of the largest ethnic group 25. In reality the correlation between these two measures across local districts is 98% and therefore this consideration is somewhat academic. Table 4 contains the results of the district level specification at the 10km scale using the population share of the largest group as the measure of district diversity. Column (1) includes country and province fixed effects, ethnic controls and average community wealth but no other controls. Column (2) adds the full slate of geographic and economic controls. Column (3) employs an instrumental 23 As mentioned above in the context of diversity instruments if more than one household member responded I use the longer reported tenure. Households also had the option to report Always and this was coded at the maximum possible tenure. The results are robust to variations in this measure. 24 Fractionalization is identical to the Herfindahl of industry concentration used in Industrial Organization. It is calculated as F = 1 E e=1 p2 e. 25 As noted above fractionalization would be the correct measure if there were no spillovers between groups and the share of the largest group would be the correct measure if spillovers between groups are complete. 19

20 variables approach to control for potential endogeneity of diversity while column (4) adds instruments for the wealth measure (while still instrumenting for diversity). The results demonstrate that moving from a completely homogeneous community to a heterogeneous community would, on average, reduce the fraction of households that access piped water by 10%. This effect seems relatively large given that just under 30% of households access piped water across the sample. This result is therefore consistent with the model and indicates that either the preferences or institutions story is relevant to this case (or possibly both). Not surprisingly wealth has a large and significant effect on access to piped water and the relatively minimal effect of incorporating instruments indicates that the measurement issue mentioned above may not be very significant 26. In addition to ethnic diversity the model suggests that there may be a significant scale effect as the efficiency of public good provision increases with the density of the population. The results indicate that population density (urbanization) offers a significant increase in the likelihood that a household will access piped water 27 The fraction of households with a female head is also associated with an increase in piped water access. While it is outside the scope of this paper to analyze this result further it is consistent with the existing literature on gender and household spending on necessities(thomas 1990). Of the geographic variables considered the only one that is significant is the distance to a river. While it is always significant the effective variation is exceedingly small as moving 10km further from a river increases the fraction of households that access piped water by 2%. From the positive effect we can infer that the increased cost of alternatives to piped water is more important than any potential cost reduction associated with a cheap water source Alternate Measures of Ethnic Diversity Column (1) of table 5 repeats the primary specification with local ethnic fractionalization as the measure of diversity. Not surprisingly given the correlation between the two measures the results are not significantly different. However in column (2) both measures are included at the same time and the results indicate that the share of the population in the largest group is the dominant effect. The positive result for fractionalization is likely related to the fact that desirable communities will attract more diverse populations and therefore after controlling for the positive effect of homogeneity 26 I continue to instrument for wealth in the vast majority of alternative specifications as it is possibly a problem and the results are not generally affected. 27 The efficiency of piped water in urban areas may also be linked to poor quality alternatives as rivers, streams or underground aquifers may be of lower quality than in rural areas. 20