Social Structure and Redistribution: Evidence from Age Vs Kin Based Organizations∗

Social Structure and Redistribution: Evidence from Age vs Kin Based Organizations∗

Jacob Moscona† and Awa Ambra Seck‡

August 9, 2021

Abstract

We document that ethnic groups’ social structure shapes patterns of economic interaction and hence the impact of public policy. Our analysis focuses on age set societies, ethnic groups comprising over 130 million people in sub-Saharan Africa in which individuals are organized into social groups based on age, known as “age sets,” that take priority over kin. Ethnographic accounts suggest that in age set societies, within-cohort economic ties are strong while inter-generational family ties are comparatively weak. First, we analyze a randomized unconditional cash transfer program in Northern Kenya and document that in age set societies, but not in kin-based societies, there are large consumption spillovers within the age cohort. Moreover, focusing on an arm of the experiment that simulates a pension program, we ﬁnd that randomly increasing the income of older people improves child health and increases household education spending in kin based societies, but has no such impact in age set societies. Next, exploiting the staggered roll-out of Uganda’s social pension program, we document a similar pattern at a national scale: household exposure to the pension program has a large, positive effect on child health in kin-based societies, but no impact in age set societies. These ﬁndings highlight how local variation in social structure can lead to markedly different, yet predictable, patterns of redistribution, thereby shaping the consequences of national policies.

∗We thank Daron Acemoglu, Alberto Alesina, Emily Breza, Rema Hanna, Michael Kremer, Nathan Nunn, Ben Olken, Gautam Rao, James A. Robinson, and Robert M. Townsend for their guidance and support. We also thank seminar partici- pants at Harvard, MIT, DEVPEC, and CEPR/LEAP for their comments. †Department of Economics, Massachusetts Institute of Technology, 50 Memorial Drive, Cambridge, MA 02142, U.S.A. (email: [email protected]; website: http://economics.mit.edu/grad/moscona) ‡Department of Economics, Harvard University, 1805 Cambridge Street, Cambridge, MA, 02138, U.S.A. (email: [email protected]; website: https://scholar.harvard.edu/awaambraseck) 1 Introduction

Informal income transfers are pervasive, particularly in low-income contexts, and the pattern and scale of these transfers shape the impact of public policy (e.g. Egger et al., 2019). When credit and insurance markets are imperfect, the ability to share risk and redistribute income is key for preventing major individual hardship (e.g. Rosenzweig and Stark, 1989; Townsend, 1994, 1995). But what determines real-world patterns of ﬁnancial ties and redistribution? Economic analysis has often presupposed that the same interests and incentives are paramount in all settings—for example, individual desire to minimize consumption risk—and that individuals are most connected with members of their family or village, social groups familiar in the West (e.g. Townsend, 1994; Kocherlakota, 1996). Policy analysis and design are informed by a similar set of as- sumptions, including the idea that patterns of re-distribution are driven by the same passions across contexts and can hence be expected to look similar.1 Anthropologists since Franz Boas(1887) and Robert Lowie(1917), however, have documented vast differences in the patterns and goal of economic exchange across cultures and argue that ostensibly clear-cut social groupings, like the family, vary widely in relevance. Local patterns of allegiance are persistent and not always free to adjust as individual or regional economic conditions change.2 The anthropologist’s perspective would be that patterns of exchange and their underlying causes are vastly different across contexts and can only be understood through the lens of local social structure. This paper empirically investigates the extent to which long-standing features of social organization predict the shape of modern economic networks and the spillover effects of public policy. Our focus is the stark distinction between societies organized around kin-based social groups and societies organized around age-based social groups, known as “age set societies.” In the latter, the primary social unit is the “age set,” a group of individuals who are a similar age and are initiated into adulthood at the same time. In age set societies, the strongest feelings of obligation and allegiance are typically between individuals of the same age cohort, rather than family or extended family. Ex- amples of age set organization exist around the world but are particularly common in sub-Saharan Africa, where we estimate at least 130 million people live in age set societies; it is described by Lowie (1921) as the “most important example of an alternative to kinship-based organization.”3 This unique setting, in which two markedly distinct forms of social organization co-exist in the same area, allow us to provide direct evidence of the relevance of local social structure for understanding modern redistribution and the spillover effects of public policy. Age set organization has been thoroughly studied by anthropologists. Their qualitative accounts

1For example, as discussed in more detail below, it is often presumed that pension programs will have positive spillover effects on children because, in kin-based societies, grandparents are expected to invest in grandchildren. Uganda’s Ministry of Gender, Labour, and Social Development uses this idea to justify its Senior Citizen Grant program; see here: https:// socialprotection.go.ug/senior-citizens-grants-scgs/. Spatial models often assume a parametric spatial structure to spillover effects that are the same throughout the sample (e.g. Egger et al., 2019) 2While a growing number of economic analyses directly measure local ﬁnancial networks (See e.g. Ambrus et al., 2014; Breza et al., 2019; Kinnan et al., 2020), many contexts pose limits on such detailed data collection, leaving us with a limited understanding of the structure and formation of social and ﬁnancial ties. We describe this work more in detail below. 3Examples include the Oromo, Zulu, Maasai, Kikuyu, Samburu, Tiv, and Iteso; while age set organization exists in West Africa (e.g. the Tiv), to our knowledge it is most prevalent in East and South Africa. The 130 million estimate combines our coding of the existence of age set organization across groups and the most recent census data from each country.

1 suggest that in age set societies, compared to kin based societies, within-cohort economic ties are strong and within-family and inter-generational ties are weak (e.g. Radcliffe-Brown, 1929; Dyson- Hudson, 1963; Morton, 1979; Handley, 2009). According to Eisenstadt(1954): “[R]elations between age mates [involve] general and permanent obligations of cooperation, solidarity, [and] help, closely resembling, in this respect, the family and kinship.” Still today, in age set societies in Kenya it is “toward age mates that men look for support rather than brothers or close patrilineal kin” (Spencer, 2014, p. 48). Age mates are initiated together at a young age and pass through different phases of life together, creating a strong within-cohort bond that persists until an individual dies.4 Anthro- pological accounts not only provide evidence that age set societies are a prevalent form of social organization throughout sub-Saharan Africa, but also that across studied examples they lead to a strikingly different structure of allegiance and pattern of resource redistribution (Morton, 1979).5 We use a wide range of ethnographic accounts to compile data on the social structure of as many ethnic groups as possible in Kenya and Uganda, the focus of our empirical analysis. Information on the presence of age set organization is not available from standard ethnographic databases for sub-Saharan Africa, including the Ethnographic Atlas. Therefore, following the canonical definition of age set organization from Radcliffe-Brown(1929), for each ethnicity we coded by hand whether the dominant social form is the age set or the kin group. By population, we were able to definitively determine the dominant form of social organization for 99% of the studied sample in Kenya and 85% of the sample in Uganda, where we find members of age set societies comprise 72% and 27% of the population respectively. Section2 and Appendix A1 describe this coding process in detail. Our main hypothesis is that age set organization affects the structure of local financial ties by en- couraging stronger within-cohort links and weaker within-family and inter-generational links. Fur- ther, we argue that this translates in different patterns of redistribution which, in turn, affect the spillover effects of national policy. The first part of our empirical analysis homes in on a randomized evaluation of Kenya’s Hunger Safety Net Program (HSNP), a large-scale cash transfer program, which to our knowledge is the only setting with randomized cash transfers and all baseline data re- quired for investigating our hypothesis.6 Sub-locations in Northern Kenya were randomly assigned to treatment and control groups. Within both treatment and control locations, individuals were selected to receive the cash transfer via one of several targeting mechanisms; in treatment locations delivery of the transfers began immediately while in control locations they were delayed for two years. Crucially, there are age set and kin based societies in both treatment and control sub-locations. We first exploit the randomized transfer experiment and a triple differences design to estimate within-cohort consumption spillovers, separately for members of age set and kin-based societies. In age set societies, randomly increasing the income of an individual’s cohort members via the transfer program dramatically increases consumption spending; consistent with the HSNP’s goal to reduce

4In Northern Kenya, age mates of all ages “share milk, meat, tea, sugar, etc” and “can rely on each other even if you have no money or social problems” (Handley, 2009). 5For example, see Baxter et al.(1979) on the Oromo; De Heusch(1985) on the Zulu; Spencer(2014) on the Maasai; Foner and Kertzer(1978) on the Kikuyu; McCluskey(2013) on the Iteso; Dyson-Hudson(1963) on the Karimojong; Bohannan (1964) on the Tiv; and Handley(2009) on the Samburu, Turkana, and Rendille. 6In particular, as will become clear, our analysis requires local variation in age set organization (in both treatment and control groups), as well as information on household-level ethnicity, age composition, and consumption. Furthermore the data was made public and is available at https://microdata.worldbank.org.

2 hunger, the effect is driven by expenditure on food. There is no evidence of these within-cohort spillover effects in societies without age sets. A one standard deviation in age cohort income increases individual consumption by 0.3 standard deviations in age sets societies and has zero detectable impact in kin-based societies. Moreover, spillover effects in age set societies are highly concentrated in an individual’s own cohort—as a falsification exercise, we document that there are no estimated spillover effects from random transfers to nearby cohorts in age set societies, consistent with the effect being driven by age set organization itself. The spillover effects are driven especially by younger and older cohorts. The strong effect for the young suggests that the relevance of age set organization is not diminishing systematically over time, while individuals at peak income earning potential seem less reliant on their age set network. We interpret these findings as evidence of strong within cohort financial redistribution in age set societies. This first set of findings suggests that age set societies have substantially stronger within-cohort consumption spillovers; however, age set ties could be either in addition to family ties or, as sug- gested by ethnographic accounts, instead of family ties. To directly investigate differences in inter- generational within-family ties, we turn to an analysis of one of the experiment’s treatment arms—a pension program. It is often hypothesized that pension programs benefit children, including through health and educational investment, because grandparents invest in their grandchildren (e.g. Bertrand et al., 2003; Duflo, 2003). However, if inter-generational ties are weaker in age set societies, pension recipients will be less likely to invest part of the pension in the younger generation, leading to a more limited (if not null) spillover effect of the program on children. We document that, for households in kin-based societies in sub-locations randomly assigned to the pension program treatment, the weight-for-height of children under five years old became significantly higher. Moreover, spending on education—measured either as absolute spending or a share of total expenditure—significantly increased. However, we find no such effects for members of age set societies; the effect of the pension program on measures of child health and education spending are both statistically insignificant and very close to zero. These findings are a first indication that, in age set societies, inter-generational family ties are substantially weaker and hence the distributional impact of social policy starkly different. The second part of our empirical analysis investigates the broader external and policy importance of understanding local variation in social structure and how it shapes financial ties. We investigate the staggered introduction of Uganda’s national Senior Citizen Grant (SCG), a pension program for individuals over 65 that was piloted beginning in 2011. Even the Ugandan government anticipated that it would have broad positive effects on children; the website for Uganda’s Ministry of Gender, Labour and Social Development reads: “Although the SCGs target older people, the grants benefit more than only the senior citizen beneficiaries[; they] have significant impact on development outcomes as older people tend to invest a portion of their grant money in meeting their grandchildren’s nutritional, health, education needs.” More broadly, pension programs are a canonical example of social policy expected to impact many individuals besides the direct beneficiaries, and for which spillover effects have been the subject of extensive analysis. To investigate the impact of the SCG on children in societies with and without age sets, we exploit the fact that the SCG was piloted in 2011 in fifteen treatment districts, combined with the fact that

3 both pilot and non-pilot districts contain members of age set and kin-based societies. Using Uganda’s 2016 Demographic and Health Survey (DHS), we estimate the “potential pension exposure” of each extended household using information on the age distribution of its members.7 In our baseline measure, this captures the number of pension-years each household would have received from 2011-2016 had their household composition remained fixed. Comparing more versus less exposed households in pilot versus non-pilot districts in a difference-in-differences framework, we estimate the effect of the pension program on child health in societies with and without age sets. While we find a large, positive effect of pension exposure on the weight and prevalence of extreme malnutrition of children in societies without age sets, we find no observable effect of pension exposure on child health in societies with age sets. The difference between the effect for societies without and with age sets is itself highly statistically significant, and the findings are robust to the inclusion of a broad range of household and child-level controls, as well as alternative parameterizations of the pension exposure measure. Finally, we find similar effects focusing on the school attendance of older children; however, these results are strongly driven by male children. Thus, the weaker inter-generational links in age set societies substantially dampens the benefits of Uganda’s pension program for children; an additional year of pension receipt in an extended household without age sets, compared to one with age sets, increases child weight by roughly 0.15 standard deviations and reduces the likelihood that a child is malnourished by roughly 5.5%. To rule out the possibility that the findings are driven by unobservables correlated with household composition or differential trends in ethnic groups with and without age sets, we present two sets of falsification tests. First, we construct a placebo pension exposure measure based on extended household members who are just below the age of pension receipt. We find no differential effect on societies with and without age sets of this placebo measure, suggesting that the effect is not driven by the selection of households with more or fewer older household members. Second, we replicate our baseline estimates except we use the 2006 round of the DHS, prior to the introduction of the pension program pilot. Again, we find no differential effect of pension exposure on societies with and without age sets. We also find qualitatively similar results using an alternative estimation strategy that does not rely on variation in household composition and only exploits the introduction of the pension pilot; these estimates suggest that the baseline findings are not driven by differential changes in household composition in response to the pension pilot in societies with and without age sets. Additional evidence presented throughout the analysis show that the markedly different structure of financial ties and policy spillovers in age set societies is not driven by other important cultural or social factors. First, we present a broad range of qualitative, ethnographic work on age set organization in Section2, that is consistent with our findings and proposed mechanism. Second, both across ethnicities in our sample and across households in each empirical setting, we do not find systematic evidence that age set organization is correlated with other observable features of culture and social or political organization, including customs related to marriage, inheritance, and gender; household

7Interestingly, we do not ﬁnd evidence of signiﬁcant differences in household composition between age set and kin- based societies in this setting (see Table A12). This is consistent with the fact that we are not aware of any mention of differences in residential sorting in age set societies in ethnographic work, or any work suggesting that individuals would actually reside with age mates or apart from family. Nevertheless, we also report results from a robustness check that does not exploit any household-level variation (see Figure A1).

4 composition; the prevalence of violent conflict; and pre-colonial economic development. This could be due to the fact that our relatively narrow geographic focus holds many features of society fixed. Finally, controlling directly for the role of ethnicity-level political and economic organization does not affect our baseline estimates; indeed, it is hard to imagine an omitted variable that would drive the unique and very specific set of hypotheses we test in the empirical analysis. We conclude with a discussion of what the different structures of economic interactions in kin- based and age set societies imply for patterns of inequality. In kin-based societies, where there are strong connections between members of different age cohorts, individuals may be better positioned to smooth consumption over the lifecycle. In age set societies, however, where these ties are weaker, individuals at stages of life when they have less income earning potential—the young and the old— might consume systematically less since their network is composed predominately of similarly aged individuals. The consumption data from the HSNP baseline data is consistent with this conjecture: while consumption is relatively flat over the lifecycle in kin-based societies, in age set societies it is hump-shaped, with younger and older individuals consuming less. Meanwhile, we find that there is more inequality across clans in kin-based societies compared to age set societies, where social networks cut across families and lineages preventing any extended family from accumulating much more than any other. These empirical patterns further highlight that social structure is an important determinant of which members of society are well connected and which are left most vulnerable. Our study of age set organization builds on a range of work documenting the importance of social structure for economic development. Most prior work has focused on features of kinship structure, including whether inheritance is patrilineal or matrilineal (e.g. La Ferrara, 2007b; La Ferrara and Milazzo, 2011; Lowes, 2016), whether the society is matrilocal or patrilocal (Bau, 2019), whether the society is organized into “segmentary lineages” (Moscona et al., 2017, 2020), and the importance of kinship tightness and family ties for long run development (Greif, 1994; Alesina and Giuliano, 2010, 2014; Akbari et al., 2017; Enke, 2019). Our focus on network structure is related recent studies of the relationship between family ties and political networks in development (e.g. Naidu et al., 2015; Cruz et al., 2017). Like previous work, our findings show that social structure shapes present day economic activity; however, moving beyond a focus on kinship, our work suggests an important role for “horizontal” social ties in economic development. Our study is also related to a large literature investigating the consequences of informal risk sharing and redistribution (e.g. Rosenzweig and Stark, 1989; Townsend, 1994, 1995; Udry, 1994; Ravallion and Chaudhuri, 1997; Fafchamps, 2003, 2011; Fafchamps and Gubert, 2007). A range of theoretical work examines the formation of financial networks, mostly focused on the desire to minimize consumption risk (e.g. Kocherlakota, 1996; Bramoullé and Kranton, 2007; Bloch et al., 2008; Ambrus and Elliott, 2020). More specifically, this paper contributes to the growing literature on the relationship between local network structure and risk sharing (see Breza et al., 2019, for a review). Ambrus et al.(2014) develop a model that shows how local obligations allow individuals to have “close to per- fect” insurance; the effectiveness of the insurance depends on the network structure. Angelucci et al. (2018) provide an indirect empirical test of this hypothesis, showing that the cash transfers provided by Mexico’s Progresa program are redistributed within extended family. Finally, Munshi and Rosen-

5 zweig(2006), can not reject full insurance at the village-by-caste level in India. 8 Remaining agnostic about the efficiency of each network structure, we show that differences in social structure lead to markedly different redistribution patterns, even among close ethnic groups subject to very similar economic conditions. An implication of this finding is that ethnographic data could be an important tool for predicting network structure where direct measurement is impossible, and thereby inform policy design. Pension programs are a key example; indeed, countries with a significant population of age set societies that have adopted large scale pension programs include Kenya, Tanzania, Zambia, South Africa, Nigeria, Namibia, and Mozambique. Last, our findings contribute to an established anthropological literature on the social role of age set organization and other forms of “cross cutting ties”— relationships that cut across families and other social groupings (see Bernardi, 1985, for a canonical discussion). A range of qualitative evidence that the prevalence of age sets shapes individuals’ pattern of social connections, and even argue that in age sets societies, age mates are essential for risk sharing and redistribution (e.g. Eisenstadt, 1954; Dyson-Hudson, 1963; Bohannan, 1964; Morton, 1979; Handley, 2009; Spencer, 2014). We document the important implications of age set organization for understanding the impact of public policy, including its effect on child health. This paper is organized as follows. The next section describes age set organization, as well as our sample, and discusses qualitative work on the role of age sets in risk sharing and redistribution. Section3 estimates income spillover effects in age set compared to kin-based using a randomized evaluation of Kenya’s Hunger Net Safety Program. Section4 analyzes the role of age sets in shaping the pattern of spillovers from Uganda’s pension program expansion. Sections5 and6 discuss additional implications of our findings and conclude.

2 Age Set Systems

2.1 Deﬁning and Identifying Age Set Organization

Age set societies are ethnic groups in which the dominant social group is the “age set,” a group of men, and sometimes women, of a similar age. Individuals in an age set are “initiated” into adulthood at the same time (and often under a common name), move through life together, and turn to each other for economic support. In the words of Radcliffe-Brown(1929, p. 21), one of the ﬁrst to develop the terminology, an age set is deﬁned as:

A recognized [group] consisting of persons (often male persons only) who are the same age [...] In Africa, at any rate in East and South Africa, an age set is normally formed of all those males who are initiated at one time. Once a person enters an age set [...] he remains a member of the same age set for the remainder of his life.

These age based groups often take priority over family or kin relationships (e.g. Eisenstadt, 1954). Age set societies are widespread. According to Lowie(1921), age sets are the “most important example of an alternative to kinship-based organization.” Well-studied examples, from all regions

8See also Morduch(2004), Mazzocco and Saini(2012) and Mobarak and Rosenzweig(2013) on risk sharing within castes in India, as well as Angelucci and De Giorgi(2009) and Kinnan and Townsend(2012) on the importance of family ties.

6 Figure 1: Maps showing the district boundaries in Uganda (left) and county boundaries in Kenya (right), along with the share of Demographic and Health Survey (DHS) households in each district that are members of age set societies; country borders are also shown in bold. Districts are color coded by quartile in the share of respondent households that are members of age set societies. For each country we use the last wave of DHS available, Kenya (2014) and Uganda (2016). of sub-Saharan Africa, include the Oromo of Ethiopia (population 27 million); the Zulu and Sotho of South Africa (population 11and 6.5 million respectively); the Maasai and Kikuyu of Kenya and Tan- zania (population 1.65 and 8.1 million respectively); the Iteso and Karimojong of Uganda (population 3.2 and 1.2 million respectively); and the Tiv of Nigeria (population 6.5 million).9 Based on our data, age set societies comprise over 130 million individuals in sub-Saharan Africa alone.

2.2 Our Sample: Age Sets in Kenya and Uganda

Our empirical analysis focuses on policy experiments in Kenya and Uganda. Thus, we determine, for the societies in these regions, whether or not age set organization is the dominant form of social organization using the deﬁnition from Radcliffe-Brown(1929). 10 Data on the presence of age set organization is not available from standard data sets, including the Ethnographic Atlas (used frequently in empirical studies of social structure, e.g. La Ferrara, 2007a; Bau, 2016; Enke, 2019; Ashraf et al., 2020); thus, throughout the analysis, we determined ourselves whether or not each ethnic group is an “age set society” on the basis of extensive secondary source analysis. In Appendix A1, we list all ethnic groups in our sample, whether or not each is an age set society, and the set of sources used to

9For references on the age set organization of the aforementioned groups, see Baxter et al.(1979) on the Oromo; De Heusch(1985) on the Zulu; Spencer(2014) on the Maasai; Foner and Kertzer(1978) on the Kikuyu; McCluskey(2013) on the Iteso; Dyson-Hudson(1963) on the Karimojong; Bohannan(1964) on the Tiv. Population estimates were taken from the most recent census data. 10Throughout, when we were unable to find sufficient information to definitively code an ethnic group as having age sets or not, we excluded members of that ethnic group from the sample.

7 make that determination. We do not include any ethnic group in our sample unless we were able to determine with certainty whether it is an age set society or not; in our analysis of Kenya’s HSNP, we were able to code our age set variable for 99% of the sample, and in our analysis of Uganda’s pension program using DHS data, we were able to code our age set variable for 85% of the sample. Figure1 documents the geographic variation of age set organization in Kenya and Uganda. The ﬁgures are constructed using individual-level data from latest round of the Demographic and Health Survey (DHS) in each country that contains both district identiﬁers and information about each respondent’s ethnicity.11 Each district is colored based on the share of surveyed households that we identify as belonging to an age set society. Members of age set societies make up 72% of the households in Kenya and 29% of the households in Uganda. While here we show district-level variation, in our empirical analysis, we use only within-district variation in ethnicity and age set organization; we display the aggregated data here only for clarity. The sources that we use to determine whether each ethnic group has age sets are historical. While the presence of age set organization in the past could no doubt have long run impacts on patterns of social interaction, we are able to systematically investigate the extent to which one age set custom has persisted today. A key feature of age set societies, as noted even by Radcliffe-Brown(1929, p. 21), is an initiation ritual; in East Africa, the initiation ceremony almost always involves circumcision and takes place at around the age of puberty (Spencer, 2014; Handley, 2009). If age set customs are still prevalent today, we would expect that in age set societies: (i) the timing of circumcision is coordinated among all boys and (ii) circumcision is more likely to take place around the time of puberty. We investigate this hypothesis directly using data on the circumcision age of a representative sample of males from Kenya and Uganda from the DHS surveys. Figure2 displays a histogram of the age of circumcision for all respondents, separately in societies with and without age sets; the mean and standard deviation of each distribution is reported at the bottom of the graph. Men in age set societies are less likely to be circumcised at birth, more likely to be circumcised during or after the age of puberty, and the standard deviation of distribution of the age of circumcision is lower.12 These features of the data exist both in rural and urban settings when we plot the distributions separately. We interpret this pattern as a reassuring check that, consistent with qualitative accounts, age set ceremonies are indeed prevalent today and discernible in our data13.

2.3 The Function and Origin of Age Sets

Qualitative Work The determinants of age set organization have not been a major emphasis of anthropological or ethnographic work, which has instead focused on describing and understanding the consequences of age sets. However, while there is no broadly accepted theory of the origin of age

11The Kenya DHS survey was conducted in 2014 and the Uganda survey was conducted in 2016. 12Additionally, in Kenya, 28% of the circumcised males belonging to age set societies were circumcised in a “ritual site” while this is true for only 15% of males from kin-based societies. In Uganda, although the share circumcised at a ritual site is higher in an age set societies, the share is very small across the board since the vast majority of boys are circumcised at a health facility. 13Notice that, even if the timing of circumcision rituals differs slightly across ethnic groups, most rituals are concentrated around the time of puberty. Hence, even if we pool all ethnicities together, we are still able to identify a peak in the frequency of circumcision around the age of puberty, consistent with the timing of circumcision rituals.

8 No Age-Sets Age-Sets .15 .15 .1 .1 .05 .05 0 0 0 10 20 30 40 50 0 10 20 30 40 50 Age at circumcision Age at circumcision Mean=11.949, Sd=7.659, N=5599 Mean=14.059, Sd=4.56, N=6786

Figure 2: Age of circumcision for all male respondents in the latest DHS Surveys from Uganda and Kenya, separately in societies without age sets (blue) and with age sets (orange).

based organization, several hypotheses have been proposed. In several contexts, age sets are used to organize men in battle; this has led LeVine and Sangree (1962) to hypothesize that age set organization began as a military innovation in response to the threat of external conflict. Consistent with this hypothesis, Zulu age set organization increased in importance during the rule of Shaka Zulu, a famous leader and military strategist who organized warriors into barracks and fighting regiments based on their age set (Eldredge, 2014). Among several East African age set societies, particularly herding societies, the role of age sets in fighting remains important; Morton(1979, p. 90) notes, “[T]he age set system represents the principal means for organizing manpower for the purpose of protecting the group against forcible intrusions from the outside and conducting offensive warfare against neighboring groups.” This fighting often involves cattle raids. Another function of age sets is the regulation of marriage and political succession. An individual’s age set, and the marriage status of members of older adjacent age sets, determines whether or not he is permitted to marry; the right to have children is also sometimes determined by one’s position in the age set system (Morton, 1979). The age set system can also regulate political succession, and determine who (i.e. which cohort) is in control of political power (Dyson-Hudson, 1963). Thus, age sets may be related to the structure of marriage and the role of women in society; the latter possibility is reinforced recent ethnographic work in Northern Kenya suggesting that the age set system remains a male-dominated institution (Handley, 2009). Age sets could also be related to political inheritance and the structure of local leadership. A final view in the anthropological literature suggests that there are no clear causal factors leading

9 to the development of age set organization; instead, age sets emerged through a complex evolutionary diffusion process likely originating in the region of the Oromo in Ethiopia (Beckingham and Huntingford, 1954; Hinew, 2012). This hypothesis follows a growing set of work arguing that social structure in general emerges in ways that can be idiosyncratic and not clearly related to pre-existing factors (Smith, 1956; Salzman, 1978; Kelly, 1983). The idea that age set institutions could have dif- fused historically across space is one advantage to our within-country focus, comparing the behavior of members of societies with and without age sets in the same small geographic area.

Correlational Evidence We examine the extent to which hypotheses about the origins and cor- relates of age set organization are born out in our sample of ethnicities from Uganda and Kenya. Using data on a range of ethnicity-level characteristics predominately from Murdock(1967), we investigate whether age set organization is correlated with characteristics proxying features of society described above, including: (i) conflict and herding, (ii) political succession, (iii) marriage and the role of women, and (iv) pre-colonial economic and political development. Table1 reports a series of ethnicity-level correlations between age set organization and a broad set of ethnicity characteristics, for the set of 29 ethnic groups of our sample, that we were able to match to the Murdock(1967) Ethnographic Atlas. We do not find evidence of systematic differences between societies with and without age set in our sample of ethnic groups from Uganda and Kenya. They appear similar across our measures of violent conflict and herding, the transfer of political power, characteristics of marriage and inheritance, the role of women, and pre-colonial development. A range of potential covariates were not included in the table because there is no variation in our sample, further evidence that our sample consists of a comparable set of ethnic groups.14 The one exception—consistent with work on the role of cross-cutting ties in limiting the centralization of political power—is that age set societies come from less centralized states. Thus, while it is unclear how this would bias our main estimates and this significant difference could be due to random chance, we are careful to control for for the role of pre-colonial state centralization in our main analysis All of our findings are robust to controlling flexibly for features of pre-colonial development. The lack of major observable differences between societies with and without age sets is also consistent with a more recent view in anthropology that social structure diffuses in idiosyncratic ways and without an obvious set of causal determinants (e.g. Beckingham and Huntingford, 1954; Kelly, 1983; Hinew, 2012).15

2.4 Age Sets and Redistribution

Numerous ethnographic studies have documented that age set organization shapes contemporary patterns of economic interaction. Qualitative accounts suggest that members of age set societies have both stronger economic ties to individuals in their age cohort and weaker ties to members of other generations, including within their own family. In one early account, Eisenstadt(1954) notes: “The relations between age mates [involve] general and permanent obligations of co-operation, solidarity,

14Characteristics with no in-sample variation include: matrolocality, sex differences in animal husbandry, inheritance rule for moveable property, matrilineality and patrilineality. 15This evolutionary model is also consistent with recent evidence presented in Moscona et al.(2020).

10 Table 1: Correlations between age set organization and other ethnicity-level characteristics

(1) (2) (3) (4) (5) (6) Variable Name Sample Age Set vs. Variable Name Sample Age Set vs. Mean No Age Set Mean No Age Set

Panel A:Conﬂict and Herding

Pastoralism dependence, 0-1 3.552 0.307 Pre-colonial Conflict 0.0690 -0.0838 (0.814) (0.0605) asinh(Conflicts), ACLED 5.339 -0.283 asinh(Conflicts), Excl. Riots, ACLED 5.199 -0.147 (0.563) (0.576) asinh(Conflicts), ACLED 4.605 0.0575 asinh(Conflicts), UCDP-GED 4.087 0.583 (0.498) (0.596)

Panel A: Transfer of Leadership

Local Leader Elected, 0/1 0.174 -0.267 Hereditary Local Leadership, 0/1 0.522 -0.333 (0.375) (0.292)

Panel C: Marriage and the Role of Women

Polygamous, 0/1 0.966 0.0419 No Cousin Marriage, 0/1 0.455 -0.282 (0.0435) (0.256) Bride Price Practiced, 0/1 0.207 0.276 Plow Used Historically, 0/1 0.0690 0.132 (0.177) (0.0908) Inheritance Rule for Land, 0/1 0.913 -0.180 Patrilocal, 0/1 0.931 0.0838 (0.136) (0.0605) Women Do Not Inherit Land, 0/1 0.826 0.270 Women Particilate Less in Ag. 0/1 0.174 -0.124 (0.167) (0.0955)

Panel D: Pre-Colonial Development

Jurisd. hierarchy (local), 1-5 1.607 -0.191 Jurisd. hierarchy (beyond local), 1-5 2.607 -0.845*** (0.220) (0.289) Settlement pattern complexity, 1-8 4.586 -0.219 (0.606) Notes: The unit of observation is an ethnic group. There are 29 ethnic groups in our sample from Uganda and Kenya, although the sample size varies slightly across specifications due to missing values in the ethnographic data. Columns 1 and 4 report the ethnicity-level characteristics. Columns 2 and 5 report the sample mean of each measure and columns 3 and 6 report the difference in the characteristic between societies with and without age sets. In each case, the denepdent variable is the reported ethnicity-level cahracteristic and the right hand side includes the age set indicator and a Kenya indicator. The ACLED conflict data are measured from 1997-2010 and the UCDP conflict data are measured from 1989-2010. The paseoralism measure is computed as in Becker (2019) and the pre-colonial conflict data are from Besley and Reynal-Querol (2014). The remaining variables are from the Ethnographic Atlas. Robust standard errors are reported in parentheses.

[and] mutual help, closely resembling, in this respect, the family and kinship.” Age sets play a central role in determining individuals’ economic relationships and sense of identity. In the words of Dyson- Hudson(1963, p. 376), whose work focuses the Karimojong cluster of ethnicities in Uganda, which form part of our sample, “An age set has members [...] who are equally uniformly identiﬁed by a single shared name, community of status, and a collective role.” The fact that members of an age set share a common identity is common to age set societies throughout sub-Saharan Africa. Morton (1979, p. 82) avers that the “highest correspondence [across East African age sets societies] exists in the manner in which egalitarian behavior among age set coevals is encouraged.” One of the best studied age set societies is the Maasai, who live in Kenya and Tanzania. Spencer

11 (2014) describes that redistributing resources with age set members is both widespread and expected. During the initiation period, members of Maasai age sets “cultivate and parade an ethos of sharing,” which is characterized by “an excessive display of ‘group indulgence,’ opposed to any suggestion of self interest” (Spencer, 2014, p. 45). Sharing of resources within the age set is not limited to young men: “[Later in life, Maasai men] share milk, meat, and the marital bed with age mates" (Spencer, 2014, p. 46). Age sets take on a similar role among several ethnic groups in Northern Kenya, the setting of the ﬁrst part of our empirical analysis. Contemporary ﬁeld notes by Carla Handley(2009) document that among these age set societies—including the Samburu, Rendille, and Turkana—within-cohort ties are strengthened and inter-generational or within-family ties are comparatively weak. Handley (2009) paraphrases one interviewee who said that:

[T]hose who are rich and those who are poor [in the age set] come together so that they may share milk, meat, tea, sugar, etc so it puts everyone as equal. [I]t is then that you can’t differentiate between rich and poor because they are covering each other. Bonding within the age set helps because the morans [age mates] can rely on each other even if you have no money or social problems.

Equality is maintained within age sets and individuals turn to age mates when they encounter negative economic shocks. Thus, in age set societies, within-cohort economic relationships are strengthened as individuals rely on members of their age cohort for support. At the same time, individuals tend to rely less on family and intergenerational relationships. Maa- sai men often do not consider individuals outside their immediate family as part of their “lineage,” as “it is toward age mates that men look for support rather than brothers or close patrilineal kin” (Spencer, 2014, p. 48). The lack of intergenerational support is noted explicitly as one of the key func- tions of Northern Kenyan age sets, which exist in part “so that the pressure is taken off the mother’s house” (Handley, 2009). As among the Maasai, kin and other community networks are notably less important; according to another interviewee from the Samburu, “The group of morans [age mates] are the ones who have more responsibilities to each other, and they are bound more to their fellow moran than they are to their brothers or sisters”(Handley, 2009, emphasis not in original). Occasionally, the relationship across cohorts and generation is not only weak, it is adversarial; Foner and Kertzer(1978) document that across age set societies, conﬂict sometimes breaks out between age sets during periods of initiations since age sets in a position of power are reluctant to relinquish their status. Together, these accounts suggest a very different structure of allegiance and support than those typically analyzed.

3 Social Structure and Redistribution: Experimental Evidence

The qualitative accounts provided the previous section suggest that, compared to kin-based societies, age set societies will have stronger ﬁnancial ties within their age cohort and weaker ﬁnancial ties across generations, even within their own family. In the remaining sections, we test these hypotheses and their economic consequences in the data. In the current section, we re-analyze a randomized evaluation of the Hunger Safety Net Program in Northern Kenya; the experimental design allows us

Figure 3: Map showing the districts in Kenya included in the Hunger Safety Net Program (HSNP) randomized evaluation. These districts are (from left): Turkana, Marsabit, Wajir, and Mandera.

to cleanly estimate both within-cohort and across-cohort spillovers of the unconditional cash transfer, in both age set and kin-based societies. In the subsequent section, we investigate the national, policy impact of these differences in networks structure by estimating relationship between social structure and the impact of Uganda’s Senior Citizen Grant (SCG) on children.

3.1 Setting

Kenya’s Hunger Safety Net Program (HSNP) is a large-scale cash transfer program in Northern Kenya. In its ﬁrst phase, which took place between 2009-2012, it reached over 69 thousand households; after 2012, it expanded to include 374 thousand households, corresponding to over 2.1 million individuals. The program’s aim is to reduce extreme poverty by means of regular and substantial cash transfers. While the program is now widespread in the country, it was ﬁrst piloted in 2009 in the regions of Mandera, Marsabit, Turkana and Wajir, as displayed in Figure3. The government of Kenya, together with Oxford Policy Management, selected 48 sub-locations to conduct a randomized policy evaluation.16 The 48 sub-locations were randomly assigned to either the treatment or control group. Households in both treated and control sub-locations were interviewed at baseline (2009) and follow-up (2011);17 eligibility to receive the cash transfer was determined based on one of three targeting mechanisms, randomized across both treatment and control groups.18 Cru-

16A sub-location is a super-set of the village. 17A second follow-up was conducted in 2012 but it suffers from severe attrition. For this reason we have decided not to include it in our analysis. 18The targeting mechanisms were community based targeting, in which the community selected households to get the transfers; dependency ratio targeting, in which households were selected based on characteristics of household health and

13 Baseline Survey Cash Transfers Distributed Follow-Up Survey

Sep 2009 - Oct 2010 Oct 2009 - Oct 2011 Nov 2010 - Nov 2011 (a) Timeline of the randomized evaluation

Treatment/Control

Eligible Non Eligible

Age Sets No Age Sets Age Sets No Age Sets

(b) Structure of the randomized evaluation

Figure 4: Timeline and structure of the randomized evaluation implemented between 2009 and 2011 by Oxford Policy Management in Northern Kenya. cially, eligible households were determined in both treatment and control groups, but only eligible households in the treatment group began receiving transfers following the baseline survey. Eligible households in the control group began receiving transfers after the ﬁrst follow-up survey. Finally, non-eligible households never received the transfer. The structure of randomization is displayed in Figure4. A crucial feature of this policy experiment is that there are members of age set and kin-based societies in both eligible and non-eligible sub-groups in both treatment and control sub-locations; it is rare for a single randomized evaluation to have such substantial ethnic heterogeneity in the participant population. This heterogeneity it possible to estimate features of the spillover effects from the cash transfer program separately for members of societies with and without age sets.

3.2 Data

The HSNP evaluation reports household-level data on income and consumption, as well as a range of individual-level demographic information on each member, including the weight and height of each child under ﬁve. Consumption expenditure is broken down by purpose; in our empirical analysis, we use estimates of total spending, spending on food, and spending on education. The survey

nutrition; and a social pension program, in which households were selected on the basis of the number of household members over the age of 55. We will abstract from the targeting mechanisms in our analysis, but we will keep into consideration the fact that different eligibility criteria might select different people based on the characteristics of the ethnic group.

14 reports the age of all household members as well as the primary language spoken by the household, which we are able to link to household ethnicity.19 The dataset also includes a broad set of baseline individual-level covariates, including health status, religion, marriage status, educational attainment, and a range of wealth measures, as well as unique village and sub-location identiﬁers. Finally, it reports treatment status at the sub-location level and eligibility status at the household level (see Figure 4), along with information on the targeting mechanism through which the transfer was assigned. Using each home language listed in the HSNP data, we determined whether or not age set organization is the dominant social form for all households in the sample. Thus was done by hand from a range of secondary source material about the social structure of ethnic groups in Northern Kenya. Appendix A1 describes the data collection and ethnic group categorization in more detail.

3.3 Empirical Analysis: Within-Cohort Ties

3.3.1 Deﬁning and Measuring Age Cohorts

Our ﬁrst goal is to estimate within-cohort spillover effects in societies with and without age sets. To do this, we construct a proxy for the share of people in an individual i’s “age set” that was eligible for the cash transfer program: . ShareAgeiaev = ∑ Eligiblei0 ∑ Eligiblei0 + NotEligiblei0 (1) i06=i i06=i j∈a×e×v× f (i) i0∈a×e×v× f (i)

where i indexes individuals, a indexes age cohorts, e indexes ethnic groups, v indexes villages and

f (i) denotes the gender. In words, ShareAgeiaev is the share of transfer eligible people in an individual’s age cohort—defined as being the same age and gender, residing in the same village, and being from the same ethnic group— excluding the individual himself.20 We include only individuals of the same gender as the respondent since in this context, age sets involve predominately males and when female age organization exists, it is distinct from the male system (Handley, 2009). Finally, when we focus on household-level consumption we assign each household the age set treatment of the main provider for the household.21 Next, we test whether in age set societies, spillover effects take place within age cohorts defined in this way. In the baseline data, we already observe that in age set societies, consumption deviations from the average consumption of the age cohort is substantially lower (see Figures A2a and A2b). The same pattern does not hold and if anything is reversed when focusing on average consumption if individuals in the same village but outside the cohort of the main provider (see Figures A2c and A2d). This is a first (albeit suggestive) indication that in age set societies, income sharing takes place

19In this context, the languages reported in the survey have a one-to-one correspondence with ethnicity, which allows us to link the two in a straightforward way. 20While for some groups in our sample, like the Turkana, we know that age sets are initiated each year, in others an age set can contain individuals who are not exactly the same age. Unfortunately, we do not have data on the exact timing of initiation ceremonies; however, in all age set societies, individuals in the same cohort are part of the same age set. Thus, we use the above single-year measure as a conservative proxy for an individual’s age set throughout the empirical analysis. 21 That is, we let ShareAgehaev = ShareAgeihaev where i ∈ h is the main provider. The results are all very similar and reported in the Appendix if we instead use the household head.

15 .4

.2309 .2

0 -.01689 -.06164 -.05602

-.2 Coefﬁcient Estimate (EffectCoefﬁcient Elgible) Cohort of Share Age Sets (Control) Age Sets (Treat) Kin (Control) Kin (Treat) Sample Bin

Figure 5: Cohort Eligibility and Consumption Spending Across Four Groups. Each bar reports the coefficient estimate of the relationship between a main provider’s household’s consumption spending and ShareAgeaev. The regression is estimated on the sub-sample noted at the bottom. From left, these are: age set societies in control sub-locations, age set societies in treatment sub-locations, non-age set societies in control sub-locations, non-age set societies in treatment sub-locations. Interviewer and targeting code by ethnicity fixed effects are included in all specifications. 95% confidence intervals are reported.

within the age cohort.

3.3.2 Cohort Eligibility and Consumption Spending

Building toward our general econometric framework, we present the simple correlation between

ShareAgeaev and (log of) consumption separately for four groups: (i) age set societies in control locations, (ii) age set societies in treatment locations, (iii) non-age set societies in control locations, (iv) non-age set societies in treatment locations. The estimates are reported in Figure5. We expect no signiﬁcant correlation between cohort eligibility and consumption in control sub- locations, (i) and (iii), where cash transfers were never delivered; this is precisely what we ﬁnd. Moreover, if our hypothesis about social structure is true, we expect large within-cohort spillovers in age set societies in treatment sub-locations (ii) but no such effect for societies without age sets (iv).

We find a large and significant correlation between ShareAgeaev and consumption for households in treated sub-locations in age set societies, but an insignificant effect that is very close to zero for households in treated sub-locations in non-age set societies.

3.3.3 General Econometric Model

Exploiting variation in the share of people eligible in one’s cohort and the randomization of cash transfer delivery to some but not other eligible populations, we estimate how an increase transfer- eligible cohort members affects consumption in societies with versus without age sets using a triple

16 difference-in-differences design. The baseline specification is: = + · · ITreat · IAgeSet + · · ITreat · INo AgeSet yiaev γ0 γ1 ShareAgeaev c(i) e(i) γ2 ShareAgeaec c(i) e(i) + · · IAgeSet + · · INo AgeSet + 0 + γ3 ShareAgeaev e(i) γ4 ShareAgeaev e(i) X Γ εiaev (2) where i indexes individuals, a age cohorts, e ethnic groups, v villages, c sub-locations, and t survey years. The dependent variable is (log of) household expenditure of the household of main provider ITreat i. ShareAgeaevt is defined as in Equation (1). c(i) is an indicator that takes value 1 if the sub-location IAgeSet was (randomly) assigned to treatment and 0 otherwise. e(i) is an indicator function that is equal INo AgeSet to 1 if e has age sets and e(i) is an indicator function that is equal to 1 if ethnicity e does not have age sets. X0 is a vector of controls which always includes all single and double interactions (excluded from Equation (2) for simplicity). For robustness, we add a series of additional controls that vary across specifications, including a vector of demographic characteristics measured at baseline; the dependent variable measured at baseline; and ethnicity-by-targeting mechanism fixed effects and age fixed effects. εiaev is the error, which we cluster at the sub-location level.

Our main coefﬁcients of interest are γ1 and γ2. These capture the effect of cash transfers to one’s cohort on consumption spending, separately for members of age set and non-age set societies. Our identiﬁcation strategy compares individuals with the same share of cohort members chosen for the transfer (ShareAgeaevt) across sub-locations randomly assigned to have the transfers delivered ITreat ITreat ( c(i) = 1) or withheld ( c(i) = 0). The main hypotheses are that γ1 > 0 and furthermore that γ1 > γ2. That is, in age set societies, resources spill over within age cohorts and do so more than in societies without age sets.

Of additional interest are γ3 and γ4, which capture differential selection of particular cohorts into HSNP eligibility in societies with and without age sets. If within-cohort consumption is more correlated in age set societies, we would expect these terms to be negative since individuals with low levels of consumption are more likely to be in cohorts with a large number of transfer-eligible members. Indeed, we ﬁnd some evidence this is the case (see Figure A3). Thus, these selection terms are important to absorb so that we compare the same cohorts across sub-locations that were randomly assigned to the treatment arm or not. Moreover, as an auxiliary hypothesis and reality check, we expect that γ3 < 0, capturing the negative selection of beneﬁciaries and its correlation with cohort members’ outcomes in age set societies.

3.3.4 Results

Baseline Estimates Table2 reports our baseline estimates of Equation (2). The dependent variable is the (log of) of total, per-capita monthly consumption. Columns 1-3 report the estimated coefficients for the full sample of main providers and columns 4-6 report same specifications restricting the sample to male main providers; in Northern Kenya, age sets are a predominately male institution and we expect the results to be driven by this subsample (Handley, 2009). Across specifications, we

ﬁnd strong evidence of within-cohort spillover effects in age set societies (γ1 > 0). The coefﬁcient increases for the male sub-sample, suggesting that the estimates are driven by male main providers

17 Table 2: Age Cohort Spillover Effects On Expenditure

Dep var: Log (1+ Per Capita Total Expenditure) (1) (2) (3) (4) (5) (6) Full Sample Males Only

Share of elig in age set * ITreat * Iageset 0.203** 0.264*** 0.274*** 0.274** 0.304*** 0.339*** (0.0957) (0.0978) (0.100) (0.107) (0.105) (0.107) Share of elig in age set * ITreat * INoageset 0.00744 -0.0105 -0.0110 0.0180 -0.0113 -0.0191 (0.0660) (0.0825) (0.0824) (0.0655) (0.0796) (0.0787) Share of elig in age set * Iageset -0.0878 -0.0964 -0.124 -0.133 (0.0688) (0.0814) (0.0786) (0.0897) Share of elig in age set * INoageset -0.0487 -0.0573 -0.0529 -0.0519 (0.0487) (0.0565) (0.0443) (0.0527)

Observations 713 707 704 603 596 593 R-squared 0.559 0.593 0.598 0.563 0.607 0.616 Interviewer FE Yes Yes Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Yes Yes Age FE No Yes Yes No Yes Yes age set X Share of age set Eligible No No Yes No No Yes Mean at baseline 7.321 7.320 7.318 7.353 7.354 7.352 P-val γ1 = γ2 0.110 0.0504 0.0445 0.0548 0.0285 0.0118 Notes: Log per capita monthly expenditures. Iageset is a variable that takes value one if the respondent belongs to an ethnicity that has age sets and 0 otherwise, INoageset is a variable that takes value one if the person belongs to an ethnicity that does not have age sets. Finally, ITreat is a dummy that takes value one if the respondent is in the treatment group and 0 otherwise, We control for economic and demographic variables such as gender, education, marriage status, occupation, religion, household size, size of the cohort, and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

(columns 4-6); this is reassuring, since in this context age sets are a predominately male institution.22 The same pattern does not hold for societies without age sets, in which we find no evidence of within cohort spillovers across specifications (γ2 = 0); the difference between the two is statistically significant, especially when we focus on the sample of male main providers. The results are similar and, if anything, more precise when we define the age set as the cohort of the household’s head rather than the main provider (Table A3).23 Our estimates suggest that a one standard deviation increase in age cohort treatment leads to an 0.3 standard deviation increase in consumption spending in societies with age sets, and zero effect in societies without age sets.

Furthermore, as hypothesized, we find that γ3 < 0, although not statistically significant. This captures the fact that in age set societies, consumption levels are highly correlated within age cohort; thus, our triple difference design that accounts for this differential selection is important. To- gether, we interpret these findings as strong evidence that in age set societies, there is substantial re-distribution within the age set. The findings are very similar and, if anything, more precise when we use food spending rather

22We do not have enough observations to estimate the regression on the sample of female main providers. Our expecta- tion is that the coefﬁcient will be smaller, or null, consistent with the fact that age sets in this context are primarily a male organization. Moreover, female main providers are a very selected sample. 57% of female main providers are widowed or divorced, vs 4% of male main providers. 23Household head are, on average, older and less likely to work than the main providers.

18 .6 .6

.4 .4

0 -.2 Spillover EffectSpillover in Set Age Societies EffectSpillover in Set Age Societies

-.4 -.2 0 5 10 15 20 0 5 10 15 20 Cohort Relative to Own (+/-) Cohort Relative to Own (+/-)

(a) Total Spending (b) Food Spending

Figure 6: Actual and Falsification Estimates. Estimates of Equation (2) with the eligibility of a series of different cohorts used to compute the right hand side variables. The x-axis reports the cohort(s) relative to an individual’s own cohort. The red point (when cohort= 0) is the spillover effect from an individual’s own cohort, our baseline estimate. The dependent variable in the analysis is (log of) total expenditure. All estimates are from the most conservative specification from Table2 (columns 3 and 6); interviewer fixed effect, targeting code by ethnicity fixed effects, age fixed effects, and age set by share of age cohort treated fixed effects, along with the baseline household-level controls, are all included. 95% confidence intervals are reported.

than total spending as the dependent variable; these estimates are reported in Table A4. Across

specifications, the difference between γ1 and γ2 is significant at the 1% level. This finding is consistent with the main purpose of the HSNP being to improve food security, and with consumption spillover effects concentrated on essential items in a population close to subsistence. All estimates are largest for young and old age cohorts, with a more limited effect on individuals aged 30-50. We estimate Equation (2) separately for a sample of main providers (i) under 30 years old, (ii) 30-50 years old, and (iii) over 50 years old; the the coefficient interest estimated on each of these samples is presented in Figure A5. First, this pattern suggests that the age set structure is not diminished among younger cohorts and will likely remain an important feature of society. Second, the pattern is consistent with the cohorts that have less income earning potential themselves (the young and old), and that are less likely to have close immediate family, relying disproportionately on the age set network.

Falsiﬁcation Tests and Sensitivity Analysis In order to make sure that we are capturing spillover effects that are restricted to one’s own age set, we estimate a series of falsiﬁcation tests. In particular, we estimate a versions of Equation (2) except we replace ShareAgeaev—the share of transfer eligible people in an individual’s own cohort—with the share of people treated in the cohorts two years above and below the individual in question, four years above and below the individual in question, six years above and below the individual in question, etc. That is, we replace ShareAgeaev with

ShareAgea+/−2,ev, ShareAgea+/−4,ev, ShareAgea+/−6,ev, etc. We hypothesize is that spillover effects for age set societies should be concentrated in or nearby one’s own cohort.

Our estimates of these spillover effects are reported in Figure6. We report our estimate of γ1 from

19 each specification. Spillover effects in age set societies are concentrated within one’s own cohort (leftmost, red coefficient); spillover effects from other cohorts within the same village are statistically insignificant and much closer to zero, especially for more distant cohorts. We report the spillover effect from all even numbered cohorts relative to the individual in question (2-20). This pattern builds our confidence in our baseline estimates. In Table A7 we report the robustness of the estimates to controlling flexibly for other household- and ethnicity-level characteristics. We matched the language groups present in our sample to the corresponding societies in the Murdock Ethnographic Atlas. Many of these characteristics are constant within our sample, which is reassuring and suggests that we are generally working with a comparable set of ethnic groups.24 We flexibly control for each ethnicity’s language group, thus homing in on within language-group variation in socials structure, as well as the measures of pre-colonial development that do vary within our sample: presence of a pre-colonial state and class stratification. Our findings are robust to the inclusion of these controls. A final potential concern is would be if the findings area driven by the community based targeting (CBT) treatment arm; while treatment is still random for this targeting mechanism, the fact that the community chooses eligibility could lead to differential selection in areas with and without age sets. In Table A9, however, we replicate all baseline results excluding sub-location in the CBT targeting mechanism and the results are very similar, suggesting that the findings are not driven by differential selection into eligibility.

Direct Gift Giving We do not have data on all financial transactions across households; however, the HSNP does ask questions about gift giving. While gifts are just one of many possible ways transfers can be made across households and consumption spillovers take place, we estimate the effect of cohort-level HSNP eligibility on both the extensive and intensive margins of gift receipt. We report our results in Table A5 using the sample of male main providers. Columns 1-3 report the effect of cohort-level HSNP eligibility on an indicator that equals one if a household received any gifts; in other words, it captures whether new financial links are established as a result of the transfer program. We find no evidence that this is the case, either in the full sample or in the sample of males main providers. Columns 4-6 report the effect of cohort-level HSNP eligibility on the total value of transfers (i.e. the intensive margin). We find positive and significant effect on the value of the transfer received in age set societies. A one standard deviation increase in the share of treated people in your cohort (∼ 44%) increases the value of transfers received in the past 3 months by 420 Kenyan Schillings, which is 11.2% of what treatment households receive over the same period of time. The effect is positive but not different from 0 and substantially smaller in magnitude in societies without age sets. Although the two coefficients are not different from each other, these results suggest that direct gift giving is one potential mechanism for the consumption spillovers.

24For example, we can not control for mode of marriage or bequest practices due to their lack of variation in our sample.

20 Full Sample Male Children Only 15 20

12.21 10 10 7.862

5 0 p =.0395 -4.937 1.033

0 -10 p =.0634

-5 -20 Age Sets No Age Sets Age Sets No Age Sets

(a) Child Weight-For-Height (Full Sample) (b) Child Weight-For-Height (Male Children)

Figure 7: Pension Grants and Child Health. Each sub-ﬁgure reports estimates of Equation (3); the left column reports γ1 and the right column γ2. The dependent variable is listed at the bottom of each graph. Standard errors are clustered by sub-location and 95% conﬁdence intervals are reported.

3.4 Empirical Analysis: Inter-Generational Transfers

3.4.1 Empirical Framework

The previous section documented substantial within-cohort spillover effects in age set societies, and no evidence of within-cohort ties in the kin-based societies in our sample. While ethnographic accounts suggest that in age set societies, inter-generational and familial links are also more limited, our results to this point are also consistent with members of age set societies sharing more income in general. In this section, we directly investigate inter-generational transfers by focusing on the pension program targeting mechanism built into the randomized experiment. We restrict our analysis to the set of eligible households in both treatment and control sub-locations that were randomly assigned to the pension grant treatment arm, and estimate the impact of pension grants on children separately for members of age set and kin-based societies. We focus on results from the following estimating equation: = + · ITreat · IAgeSet + · ITreat · INo AgeSet + 0 + yi γ0 γ1 c(i) e(i) γ2 c(i) e(i) X Γ εi (3)

ITreat IAgeSet where i indexes either children or households (depending on the speciﬁcation); c(i) and e(i) continue to represent treated sub-location and age set society indicators; and X0 is a vector of baseline controls identical to those used in the previous section. Our main dependent variables are the weight- for-height of all children under ﬁve and household-level education spending; both capture changes in investment in younger generations.25 If age set organization reduces inter-generational ties that are present in kin-based societies, we hypothesize that γ2 > 0 and moreover that γ2 > γ1; if inter- generational ties are fully absent in age set societies, we further hypothesize that γ1 = 0.

25We focus on weight-for-height rather than any anthropometric measures normalized by age because the age of children in months was not recorded; below, in our analysis of Uganda’s pension program where we have more comprehensive measures of anthropometry, we document the robustness of our ﬁndings across a range of possible parameterizations.

21 log Education Spending log Share Education Spending 2 2

.997 1 .9901 1

.1352 .05033 0 0 p =.139 p =.0784

-1 -1 Age Sets No Age Sets Age Sets No Age Sets

(a) log Education Spending (b) log Share Education Spending

Figure 8: Pension Grants and Education Spending. Each sub-ﬁgure reports estimates of Equation (3); the left column reports γ1 and the right column γ2. The dependent variable is listed at the bottom of each graph. Standard errors are clustered by sub-location and 95% conﬁdence intervals are reported.

3.4.2 Results

Figure7 reports our baseline estimates of (3). In Figure 7a, the dependent variable each child’s weight-for-height percentile.26 We find no effect on child nourishment of the pension program among age set societies, but a positive and significant effect in societies without age sets. Consistent with earlier work documenting that pension grants in kin-based societies are invested disproportionately in boys, the effect is more pronounced when estimated on the sample of male children (Figure 7b). In both cases, the difference between the effect in societies with and without age sets is statistically significant (p = 0.0634 and 0.0395 respectively), although less precise than the previous section since the pension analysis requires restricting the sample to one third of the households involved in the experiment. Next, we turn to the impact of the pension program on education spending, our second proxy for investment in younger generations. In Figures 8a and 8b the dependent variables are household-level (log of) education spending and (log of) education spending as a share of total spending respectively. In both cases, we again find zero effect of the pension program in age set societies but a positive and significant effect in kin-based societies. Table A11 reports additional specifications, documenting that the distinction between age and kin-based societies is even starker focusing on male main providers, and that this pattern is restricted entirely to education spending and is not apparent in any other spending category. Together, these estimates suggest that older generations invest in younger generations in kin- based societies but not in age set societies, where resources are transferred and re-distributed within the age cohort. The next section investigates the national policy relevance of this pattern by analyzing the staggered introduction of Uganda’s national pension program and its spillover effects.

26Estimates of the relationship between the pension program and child nutrition are reported in table form in Table A10.

Figure 9: The map shows Uganda district boundaries and the set of pilot districts in the 2011 pilot program of the Senior Citizen Grant (SCG).

4 Social Structure and National Policy: Uganda’s Pension Program

4.1 Setting

In 2011, Uganda introduced the Senior Citizen Grant (SCG), a social pension program targeting all individuals aged 65 and older of 60 and older in the Karamoja region. The pension program consists of $7.50 USD monthly transfers, which corresponds to around 20% of average monthly household consumption. Rather than introduce the program nationally, the government of Uganda piloted the program in just 15 districts; the set of pilot districts are displayed in Figure9. This partial roll out provides us with both treatment and control districts, making it possible to estimate the consequences of the program. An explicit goal of the pension program was to improve the well-being of children, since the government of Uganda believed that pension recipients would invest part of the transfer in grandchildren. The website of Uganda’s Ministry of Gender, Labour and Social Development, which was in charge of implementing the pension program, reads:

Although the SCGs target older people, the grants benefit more than only the senior citizen beneficiaries. The SCGs have significant impact on development outcomes as older people tend to invest a portion of their grant money in meeting their grandchildren’s nutritional, health, education needs.27

The staggered introduction of the pension program, combined with the widespread prevalence of age set organization in Uganda and fact that inter-generational transfers were expected to be a key

27See here: https://socialprotection.go.ug/senior-citizens-grants-scgs/. Accessed March 30, 2021.

23 part of pension program’s impact, makes it an ideal setting to study how socials structure shapes inter-generational economic ties and the spillover effect of major national policy.

4.2 Data

The key data set for this analysis is Uganda’s Demographic and Health Survey (DHS). Our main analysis relies on the 2016 round of the survey; we also use the 2006 round of the DHS, which was conducted prior to the introduction of the SCG, to conduct a placebo exercise. For each household in the survey, the DHS reports information on household composition, including the age of all members. The survey also reports anthropometric data for all children under five years of age and schooling information for all school aged children. All household members selected for the individual component of the survey report their ethnicity, which we match to the household data using unique household identifiers.28 Using the survey information on self-reported ethnicity and a range of anthropological and ethnographic accounts, we determined whether age sets are the dominant social organization or not for as many ethnic groups as possible; the details of this data collection process are discussed in Appendix A1. We were able to definitively determine whether or not 85% of the households in the sample belong to an age set society. For the remaining 15%, either ethnicity was not specified, or there was insufficient secondary source information to make this determination with certainty.29 Members of age set societies comprise 29% of our sample. Finally, we determined from the Ministry of Gender, Labour and Social Development website which districts were included in the pilot program for the Senior Citizen Grant (SCG). Pilot districts are displayed in Figure9.

4.3 Empirical Strategy

4.3.1 Measuring Potential Pension Exposure

We construct a measure of “potential pension exposure” for all households in the sample, both in pilot and non-pilot districts. Our baseline potential exposure measure captures the total number of years of pension grants each extended household would have received since 2011 if they were in a pilot district and their 2016 household composition were ﬁxed:

2016 = IAge65+ PotentialExposureh ∑ ∑ iht (4) t=2011 i∈h

Age65+ where i indexes individuals, h indexes households, and t indexes years since 2011. Iiht is an indicator that equals one if individual i in household h was at least 65 years old in year t; we construct this indicator for all individuals in every household using household member age data from the

28In the 2006 survey, only language is reported and not ethnicity. This is a coarser measure of ethnic afﬁliation than the ethnic group, but we nevertheless categorize language groups into those with and without age sets and conduct our analysis at that level. 29Most un-coded ethnic groups are small and comprise less than 2% of the population.

24 4

2.383 2

.1426 0 -.4865 -.6757

-2

-4 Age Sets (Not Pilot) Age Sets (Pilot) No Age Sets (Not Pilot) No Age Sets (Pilot)

Figure 10: Pension Exposure and Child Nutrition Across Four Groups. Each bar reports the coefficient estimate of the relationship between a child’s weight-for-height percentile and PensionExposureh. The regression is estimated on the sub-sample noted at the bottom. From left, these are: age set societies in non-pilot districts, age set societies in pilot districts, non-age set societies in non-pilot districts, non-age set societies in pilot districts. District-by-ethnicity and interview month fixed effects are included in all specifications. 95% confidence intervals are reported.

DHS. Thus, this measure captures the number of years of pension grants each household would have received if its household composition observed in 2016 were ﬁxed for the preceding years.

4.3.2 Pension Exposure and Child Nutrition

As a ﬁrst validation of our hypothesis, analogous to Figure5, we estimate the simple partial correlation between each child’s weight-for-height percentile and PotentialExposureh separately in four sub-groups: (i) age set societies in non-pilot districts, (ii) age set societies in pilot districts, (iii) non- age set societies in non-pilot districts, (iv) non-age set societies in pilot pilot districts. These estimates are reported in Figure 10.

We expect no significant effect of PotentialExposureh on child nutrition in districts without the pension pilot, (i) and (iii), since they did not receive any transfers. Based on our findings from the HSNP experiment, we further expect no effect of pension exposure in age set societies in pilot districts (ii), since intergenerational ties are weak, but we expect a positive effect in kin-based societies in pilot districts (iv), where grandparents are more likely to invest part of their pension receipt in grandchildren. This is precisely what our estimates suggest. The first three coefficient estimates are small and statistically insignificant, while the fourth is positive and significant, suggesting that in societies without age sets each additional year of pension exposure is correlated with an 2.4 percentile increase in child weight-for-height. The next section describes our more general econometric framework in which we estimate the differential effect of pension exposure in societies with versus societies without age sets on the full

25 sample of children in the DHS.

4.3.3 General Econometric Model

In order to estimate the causal effect of the pension program on child outcomes separately for members of societies with and without age sets, we use the following estimating equation: = · · IPilot · INoAgeSet + · · IPilot · IAgeSet + + 0 + yih β1 PotExph d(h) e(h) β2 PotExph d(h) e(h) αed X Γ eih (5)

where h continues to index households and i indexes individuals. In the case of the main results, the sample is restricted to all children under 60 months of age (i.e. the sample for which anthropometric data were collected). d(h) is the district in which the household is located and e(h) is the household’s

ethnic group. PotExph is the household-level potential pension exposure measure described in Sec- IPilot tion 4.3.2. d(h) is an indicator that equals one if a household is located in a pension pilot district, and IAgeSet INoAgeSet e(h) and e(h) are indicators that equal one if a household is part of an ethnic group with and without age sets respectively. Throughout, standard errors are clustered by district.

Our hypotheses are that β1 > 0 and that β1 > β2. That is, we hypothesize that transfers to older household members beneﬁt children in kin-based societies—where inter-generational ties are present—and does so more than in societies with age sets—where inter-generational ties are weak.

Our estimate of interest is β2 − β1—the differential effect of the pension program on child health in societies without versus with age sets. Further, if inter-generational ties in age set societies are completely absent, we would expect to ﬁnd that β2 = 0.

Identification and Balance Our key identification assumption is that absent the pension program, differences between children in societies with and without age set societies should be unrelated to pension exposure. To investigate the plausibility of this assumption, we investigate differences between households form societies with and without age sets in the non-pilot group. Results from this analysis are reported in Table A12 and there are two key conclusions. First, there is no correlation between age set organization and measures of potential pension exposure and household composition. These variables are all measured at the household level (Panel A). Second, there is no relationship between age set organization and our measures of child nutrition (Panel B). Average child age is also balanced across both groups, which suggest that children from the two groups are not differen- tially selected into our sample due to different rates of infant mortality. These findings suggest that it is unlikely that our results are driven by baseline differences between societies with and without age sets in either child health or household composition—which could affect selection into pension receipt—between households from these two sets of societies.

4.4 Results

Baseline Estimates Our main ﬁnding is that household exposure to the pension program improved child health in societies without age sets, but had no impact on child health in societies with age sets. Interpreted through the lens of our empirical framework, this suggests that inter-generational ties

26 Table 3: Effects of Pension Program on Child Nutrition in Societies With and Without Age Sets

Dep Var = Weight for Height Percentile (1) (2) (3) (4) (5) Full Sample Male Female

Potential Exposure * IPilot* INoageset 2.477*** 2.943*** 3.135*** 5.729*** 2.748** (0.605) (1.008) (1.098) (1.428) (1.195) Potential Exposure * IPilot* Iageset -0.649 -0.818 -0.933 -2.440 0.107 (0.965) (1.026) (0.951) (2.542) (1.173)

Observations 4,112 4,107 4,107 2,020 1,976 R-squared 0.130 0.131 0.203 0.240 0.241 District x Ethnicity FE Yes Yes Yes Yes Yes Interview Month FE Yes Yes Yes Yes Yes Age Set x Potential Exposure FE No Yes Yes Yes Yes Age in months x Gender FE No No Yes Yes Yes P-val β = γ 0.006 0.009 0.007 0.008 0.122 Mean Non-Pilot 49.054 49.054 49.054 49.054 49.054 Note: The table reports the estimated coefﬁcients from equation5. The unit of observation is the child. The sample includes all children in the household who are less than 60 months of age. The dependent variable is the weight for height percentile. Potential exposure is a measure of exposure to the pension constructed as indicated in equation6. IPilot is a dummy that takes value one if the household is in a pilot district, and 0 otherwise. Iageset is a variable that takes value one if the household belongs to an ethnicity that has age sets and 0 otherwise. Iageset is a dummy that takes value one if the household belongs to an ethnicity that doesn’t have age sets and 0 otherwise. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1

in age set societies are sufficiently weak that an income shock to the grandparent generation has no discernible impact on child well-being in age set societies. Our baseline estimates of Equation (5) are reported in Table3. Following Duflo(2003), we use child weight-for-height as our baseline anthropometric dependent variable. In column 1, we control for only ethnicity-by-district and interview month fixed effects. β1 is positive and significant, suggesting that household pension exposure improved child nutrition in kin-based societies. However,

β2 is significantly smaller than β1 and is itself small in magnitude and statistically indistinguishable from zero. In the remaining columns, we add an additional set of controls and fixed effects and the estimates remain very similar. In column 2, we include pension exposure fixed effects interacted with the age set indicator, thus flexibly absorbing any differences between societies with and without age sets in potential pension exposure. In column 3, we further add age-by-gender fixed effects. In both cases, the results are similar and if anything, our estimates increase in magnitude. Finally, in columns 4-5, we investigate the effect for male and female children separately. While we find qualitatively similar results for both, the magnitude is larger for male children, consistent with existing evidence documenting disproportionate investment of family resources in boys (Duflo, 2003;

Milazzo, 2014; Barcellos et al., 2014). Our estimates of β2 − β1—the key object of interest—from the speciﬁcations in columns 3-5 are reported in Figure 11, and are statistically signiﬁcant (p<0.001) for

27 15

8.169

5 4.067

2.641

Full Sample Boys Girls

Figure 11: Baseline Estimates. Each column reports the differential effect of pension exposure on households from societies without vs. with age sets (β − γ). Moving from left to right, the sample included in each specification includes (i) all children under six, (ii) male children under six, and (iii) female children six. The dependent variable is the child’s weight-for-height percentile. Standard errors are clustered by district and 95% confidence intervals are reported. the full and all-male samples. In Table A13 we investigate the impact of the pension program on a range of additional anthropometric measures. For comparison, column 1 reports the specification in column 3 of Table3 (the most conservative specification). First, we document that the result is very similar using the weight- for-height z-score (instead of percentile) as the dependent variable (column 2). Second, we show similar results when the dependent variable is an indicator that equals one if the child is in below the 5th percentile in the weight-for-height distribution, a common cut-off used by the CDC to define malnourishment (Mei et al., 2008; Philips and Shulman, 2018). Thus, age set organization affects not only the impact of the pension program on where a given child falls in the full health distribution but also its impact on the probability that they are extremely malnourished. Third, we document that all findings presented thus far are robust using the weight-for-age distribution rather than the weight- for-height distribution (columns 4-6). Fourth, we show that the results are qualitatively similar but smaller in magnitude and statistically imprecise using the height-for-age distribution (columns 7-9). Measurable changes in child height are only affected by persistent and prolonged patterns of consumption (Hoddinott et al., 2013; Jayachandran and Pande, 2017); the average number of years of pension exposure in our sample is 1.9.30 Finally, we construct an index incorporating all nine aforementioned measures of child nutrition by computing their first principal component. Intuitively, the first principal component loads positively on all percentile and z-score measures and negatively on all indicators for being in the bottom 5% of the distribution (see Table A15). Using this index as the

30Our findings are consistent with Cahyadi et al.(2018), who find no effect of cash transfers on stunting after two years of a conditional cash transfer program, where stunting is defined as being 2 standard deviations less than the WHO’s height-for-age standard. Indeed, most analyses of cash transfer programs find no evidence of an impact on child height, especially in the short term (Baird et al., 2016; Cahyadi et al., 2018, p. 17).

28 dependent variable, we ﬁnd a strong, positive effect of pension exposure in societies without age sets and no effect on societies with age sets (difference p-value = 0.0137). On average, an additional year of pension receipt in a household without age sets, compared to one with age sets, increases child weight by roughly 0.15 standard deviations and reduces the likelihood that a child is malnourished by roughly 5.5%.31 The large, positive effect of the pension program on child nutrition in kin-based societies in our sample is completely absent in age set societies.

Falsiﬁcation Tests To make sure that our estimates indeed capture the causal effect of the pension program, we report two sets of placebo estimates. First, we want to make sure that the effect is not driven simply by unobservables correlated with having more older members in the extended household, and not the effect of the pension program. To do this, we document that there is no effect if we measure potential exposure using the number of individual just below the age cut off for the pension program. We compute this variable analogously to our baseline measure of potential pension exposure (Equation4), except rather than count household members over age 65 in year t, we count household members who would have been aged 55-64 in year t. Our estimate of the differential effect

of this “placebo” exposure measure on child health in societies without versus with age sets (β2 − β1) is reported in the leftmost column of Figure 12a and 12b. In Figure 12a, the dependent variable is the child’s weight for height percentile and in 12b it is an indicator if the child is in the bottom 5% of the distribution. In both cases, we find no significant effect of our placebo measure and, if anything, the sign is the opposite of our baseline estimates. Our second placebo check documents that our results are not driven by differential nutrition trends in societies with and without age sets in pilot vs non-pilot districts. We estimate Equation (5), defining pension exposure exactly as in Equation (4), but using data from the 2006 round of the DHS, prior to the introduction of the pension pilot.32 We would not expect to find any effect of our main independent variables during years prior to the pension program. The differential effect of pension exposure in 2006 on child health in societies without versus with age sets is reported in the center column of Figures 12a and 12b. In both cases we find no significant effect of this second placebo measure and again, if anything, the sign is the opposite of our baseline estimates. Finally, for comparison, the rightmost column of Figures 12a and 12b reports our baseline (actual) estimate of the differential effect of pension exposure on child health in societies with versus without age sets.

Alternative Pension Exposure Measures One possible shortcoming of our baseline measure is that that it uses the 2016 household age distribution to retroactively estimate pension receipt during preceding years. To document that this feature of variable construction does not drive the results, we

31These results are in line estimates from the literature. Duflo(2003) finds that South Africa’s pension program increased weight-for-height of girls by 1.19 standard deviations. The pension program she analyzed entailed monthly transfers which corresponded to approximately 248% of average income, and she estimated the effect of receiving approximately one year of pensions. If effects on weight-for-height are linear, then our results are perfectly in line with her findings. 32One downside to this analysis using the 2006 round of the DHS is that this round does not ask household members for their ethnicity; instead, it asks for the language they speak at home. While it is straightforward to link languages to ethnic groups, the language information is more aggregated and so we are left with only six languages of which one corresponds to societies with age sets. This feature of the data could also explain the relative imprecision of these placebo estimates.

29 Weight-For-Height Percentile Weight-For-Height Bottom 5% Indicator 7 .1

4.067 .05 3

.02665 1 .01523 0 -.2816 -1 -1.856 -3 -.05 -.05789 -5

-7 -.1 Exposure Below Cut-Off 2006 Estimate Actual Estimate Exposure Below Cut-Off 2006 Estimate Actual Estimate

(a) Differential Effect: WFH Percentile (b) Differential Effect: WFH Bottom 5% Indicator

Figure 12: Falsification Tests. The left column of each figure reports the differential effect of having members aged 55-65 on households from societies without vs. with age sets. The center column of each figure reports the differential effect of pension exposure measured in 2006 on households from societies without vs. with age sets. For comparison, the right column of each figure reports the differential effect of actual pension exposure on households from societies without vs. with age sets. In Figure 12a the dependent variable is the child weight- for-height percentile and in Figure 12b it is an indicator if the child falls in the bottom 5% of the distribution. Coefficient estimates and 95% confidence intervals are reported. estimate a second exposure measure only reslies on the age distribution in 2016. We define:

2016 = IAge65+ PotentialExposureh ∑ ih,2016 (6) i∈h

Estimates using this variable, reported in Table A14, capture only the impact of pension exposure in 2016, which we can measure directly. The results are very similar. Another potential shortcoming—which is likely to bias our baseline results downward—is that, for simplicity, our baseline exposure measure does not exploit variation in child age. We estimate a separate exposure measure that does take variation in child age into account; in particular, we deﬁne:

2016 h i ChildAge = IAge65+ IBorn After t PotentialExposureh ∑ ∑ iht ∑ iht (7) t=2011 i∈h i∈h;i under 5

Estimates using this alternative pension exposure measure are also reported in Table A14 and again, the results are very similar.

Alternative Mechanisms and Additional Controls Next, we explore alternative potential mechanisms, as well as the robustness of our findings to the inclusion of a broad range of controls. These results are presented in Table A16. One possibility is that the findings are driven by differences in the number of children across households. To address this, in column 1 we include fixed effects for the number of children under five in each household and the results remain the same. Second, it is possible that differences in wealth between societies with and without age sets might affect the extent

30 to which they invest financial windfalls in children. In column 2, we therefore control directly for a wide set of measures of household asset holding.33 In column 3, we control for fixed effects of the 5-point wealth index computed by the DHS. In both cases, the results are again very similar. While in Table1 we found little evidence of systematic differences between societies with and without age sets across a range of observable characteristics, we next show that the results are robust to controlling for other important features of social and political organization. To account for · IPilot · an ethnicity-level characteristic Ze(h), we include controls of the form PotExph d(h) Ze(h); we also include all relevant double interactions and direct effects. In column 4 we control flexibly in this way for fixed effects in the number of levels of jurisdictional hierarchy beyond the local community and settlement pattern complexity, both measured from the Ethnographic Atlas; these are frequently used measures of historical political and economic development respectively.34 In column 5, we add controls for language group fixed effects, thus flexibly absorbing differences in each household’s coarser language category. The estimates remain very similar. An additional possibility is that the results are driven in part by the fact that different interviewers surveyed different ethnic groups, and that this introduces bias in survey measurement. To address this, in column 6 we include a full set of interviewer fixed effects and the estimates are again similar. Finally, in column 7 we include all controls from columns 1-6 in a single specification. Together, these results build confidence that our findings are driven by the effect of age set organization on inter-generational transfers, and not some omitted household or ethnicity-level characteristic.

Household Composition Endogeneity Our main analysis exploits differences across extended households in the number of pension recipients. Our triple-difference estimates allow for the inclusion of ethnicity-by-district fixed effects, which we view as important for isolating the effect of pension exposure on households with and without age sets. However, a potential shortcoming of our baseline specification (Equation5) is that it could be potentially biased if the introduction of the pension program induces differential endogenous changes in household composition between members of societies with and without age sets. There are several reasons this should not be a major concern. First, we find that households with and without age sets are balanced across a range of observable measures of household composition in non-pilot districts (see Table A12). Second, we control directly for pension exposure by age set organization fixed effects to flexibly absorb any differences in household composition across ethnic groups. Finally, based on ethnographic accounts, it seems highly unlikely that the pattern of selection would go in the direction of driving our results. For our estimates to be driven by selection, it would have to be that the pilot program induces grandparents who might feel less attached to their grandchildren in age set societies compared to societies without age sets to begin co-residing with grandchildren. If anything, since the default in age set societies is that inter-generational ties are weaker, we would expect the pattern of selection to be the opposite. Nevertheless, to investigate whether our findings are driven by our measured variation in house-

33Household asset controls include indicators for the presence of electricity, radio, television, a refrigerator, a bicycle, a motorcycle, and a car or truck, as well as fixed effects for main floor material, main roof material, main wall material, the type of toilet facility, and the number of rooms used for sleeping. 34In order to have as large a sample as possible, we also include a fixed effect for each variable that indicates when the value is missing in the Ethnographic Atlas.

31 Primary School Attendance

.05

.02978

.01008 0

-.01179

-.05

Full Sample Boys Girls

Figure 13: Primary School Attendance. Each column reports the differential effect of pension exposure on households from societies without vs. with age sets. Moving from left to right, the sample included in each specification includes (i) all children 6-14, (ii) male children 6-14, and (iii) female children 6-14. Coefficient estimates and 95% confidence intervals are reported. hold composition, we present estimates from an alternative specification that only exploits district and ethnicity-level variation, and hence does not depend on household composition: = · IPilot · INoAgeSet + · IPilot · IAgeSet + + 0 + yih β1 d(h) e(h) β2 d(h) e(h) αe X Γ eih (8)

Our hypothesis now is simply that in districts that receive the pension pilot, on average children from societies without age sets should be relatively better off. While this specification exploits much coarser variation than our baseline estimates, qualitatively similar results would be reassuring. Our estimates of β2 − β1 are reported in Figure A1. Our double-difference results are qualitatively very similar to our baseline estimates. whether child nutrition is measured as the weigh-for-height percentile or as an indicator that equals one if the child is below the fifth percentile. On average, compared to children in age set societies, children in societies without age sets are relatively higher in the weight-for-height distribution and relatively less likely to be malnourished where the pension pilot was introduced. These estimates suggest that the baseline findings are not driven by endogenous shifts in household composition that differs between societies with and without age sets.

4.4.1 Additional Outcomes: Older Children

We next investigate the impact of pension exposure on the school attendance of primary school age children. While we expect a smaller effect here since there is a more tenuous link between family investment and child school attendance than nutrition, Figure 13 reports the differential effect of pension exposure on the primacy school attendance of children from societies without vs. societies with age sets (β2 − β1). The ﬁndings mirror our results on the nutrition of young children, except here the ﬁndings are strongly driven by male children (middle column). An additional year of pension

32 Secondary School Attendance Married .08 .1

.06 .05

.02854 .04 0 .035 -.01361 .02329 -.05 .02 -.05925

.002664 -.1 0

Full Sample Boys Girls Full Sample Boys Girls

(a) Secondary School Attendance (b) Marriage

Figure 14: Older Children. Each column reports the differential effect of pension exposure on households from societies without vs. with age sets. In 14a, the dependent variable is secondary school attendance, and in Figure 14b, it is marriage. Moving from left to right, the sample included in each specification includes (i) all children 15-18, (ii) male children 15-18, and (iii) female children 15-18. Coefficient estimates and 95% confidence intervals are reported. receipt increases the likelihood that a primary school age boy is in school by roughly 3% in societies without age sets, relative to societies with age sets. Table A17 reports the full set of regression estimates and shows that the effect is driven by a positive and significant effect of pension exposure on school attendance of boys in societies without age sets, and an effect of pension exposure on school attendance of boys in societies with age sets that is statistically indistinguishable from zero. However, pension receipt does not affect the school attendance of girls in either set of ethnic groups. Given that in our sample most of the ethnic groups are patrilocal, and none are matrilocal, this finding is consistent with Bau(2019) who, in Ghana and Indonesia, finds that grandparents tend to invest in the education of children that will co-reside with them as adults. Finally, we turn to the outcomes of older children, aged 15-18: secondary school attendance and marriage. Our estimates for these two outcomes are displayed in Figure 14, structured in the same way as Figure 13. We continue to find a positive effect on school attendance for boys, although it is statistically imprecise on this older sample (Figure 14a, middle column). However, strikingly, we find a significant negative effect on school attendance for girls in kin-based compared to age-based societeis (Figure 14a, right column). The decline in girls’ schooling can largely be accounted for by an increase in marriage—girls from households that receive pension grams are more likely to be married kin-based compared to age-based societies (Figure 14b). This finding is driven by a large, positive effect of pension exposure on girls’ marriage in societies without age sets, and zero effect in societies with age sets (Table A18). Marriage in rural Uganda is often arranged by a girl’s household and extended family (Green et al., 2009, p. 5). For the family, there is a trade off between the social value of having married daughters and the potential economic costs of losing household labor (Bantebya et al., 2014). Early marriage also serves to prevent any stigma associated with a potential pre-marital pregnancy and, according to many parents, limit the likelihood of HIV exposure (Green et al., 2009,

33 ihu g es l n on ebr faestsceis hs ewr scmoe flow- of composed is network whose societies, set age and of with members societies young for and separately Old groups, age sets. across age without evaluation randomized HSNP the in providers could lifecycle the more over inequality are lower. children, individuals or be where grandparents, of however, parents, higher-earning societies, predominately on kin-based composed rely In to is likely stage. network lifecycle economic lifecycle same their the the since of at off parts individuals low-earning worse at systematically individuals left support, be economic economic for could of cohort pattern age and own their extent suggests on the societies ately of set determinant age important to an compared be inequality kin-based also in could exist structure that social ties that The economic groups. of ethnic across set and ways different different shocks vastly income starkly in networks; economy local financial the shapes through structure propagate policy social public local that show findings paper’s This Discussion the of impact overall 5 the on effect profound a had difference This groups. ethnic evidence societies. across no program set pension but age societies of in kin-based facilitation ties in the such ties in of inter-generational or of education, evidence nutrition, strong introduced—in find was marriage—we program pension household. the the when in ations living daughters marry to family the frees resources of influx the grants, pension age members. 12). (orange) household v, of with pp. number and the (blue) for control without we societies and of expenditure total members of) for (log separately measures y-axis (y-axis), The provider sets. main the of age 15: Figure 35 .5). p. 2009, al. , et (Green popular widely was daughters for marriage early parents, surveyed Among rprscnupinepniuefrml main male for expenditure consumption reports 15 Figure pattern: this of evidence some find We gener- younger in make might generations older that investment of forms measurable all Across 35 ntecross-section the in osmto Deviations. Consumption hs tapasta hnodrmmeso oshlsi i-ae oite receive societies kin-based a in households of members older when that appears it Thus,

Log(1+Tot Exp) 8.75 8.8 8.85 8.9 8.95 20 naestsceis o xml,sneidvdasrl disproportion- rely individuals since example, for societies, set age In . isatrposo oshl osmto xedtr yai)v.the vs. (y-axis) expenditure consumption household of plots Binscatter 40 No age-sets Age ofthemainprovider 34 Age-sets 60 80 earning individuals on average, consume systematically less; meanwhile, the evolution of consumption over the life cycle for kin-based societies, where inter-generational links are common, is comparatively flat. This difference in consumption patterns is statistically significant, as documented in Table A8, and is not driven by differences in the evolution of income over the lifecycle in societies with and without age sets.36 A very different pattern may exist across family lineages or clans. In kin-based societies, where resources are kept to a larger extent within the extended family, certain clans may accumulate substantially more wealth and income opportunities than others. In age set societies, where age sets cut across the lineage structure, this form of inequality could be more limited. Consistent with this conjecture, the standard deviation of household consumption across clans in the baseline HSNP data is 50% larger in kin-based societies societies compared to age set societies; the standard deviation of consumption across sub-clans is 60% larger. The structure of financial ties in age set societies thus might make it less likely that individual lineages accumulate a disproportionate share of local wealth, while others are left poorer and without access to financial resources. While only suggestive, this collection of evidence that social structure may predictably shape wealth and income inequality, providing a new window into analyses of optimal policy, policy targeting, and local political and economic development. There are a range of other consequences of differences in local network structure that are beyond the scope of this paper but potentially exciting areas for future work. We focus on financial networks, but age set and kin-based societies also may have very differently structured networks of information. In the former case, individuals may learn from and share more information with their peers, rather than relatives or older individuals. These patterns could have an impact on the rate and extent of technology adoption, and the spread of information more generally. Finally, the differences in social structure documented in this paper could have an impact on patterns of migration. On the one hand, in kin based societies young migrants may be more able to borrow or access information about past migrations from older individuals. On the other hand, however, the age set could provide an important network for navigating in the migration destination and make it possible for individuals to migrate in larger groups. Methodologically, our results suggest that a deeper understanding of social structure, and anthropological analyses thereof, could be a useful complement to direct network measurement and analysis. Using our reading of existing ethnographic work to identify age set organization, we were able to predict where particular social ties would be strong and weak, and hence how both Kenya’s cash transfer program and Uganda’s pension program would propagate through the economy. The approximation that our method provides could be useful for policy analysis and design, especially for large-scale programs or more remote areas where network data collection is prohibitively costly or impossible. Age set organization is one of many examples of local variation in social structure that could shape the pattern of financial, informational, or migration networks.

36 Figure A4 reports the same result as Figure 15 after controlling for log income and the pattern is virtually unchanged.

35 6 Conclusion

Vast differences in social structure exist around the world; patterns of allegiance vary widely across contexts. Our analysis documents that social structure predictably shapes patterns of economic interaction, redistribution, and the impact of public policy. We focus on age set societies, ethnicities in which individuals are organized in to age-based social groups; in these ethnic groups, an individual’s relationship with his “age mates” are of upmost importance and often take priority over family ties. Despite its prevalence in sub-Saharan Africa and around the world, age organization has not been studied by economists and suggests a strikingly different pattern of social and economic interaction than those typically modeled. Exploiting unique features of a cash transfer experiment in Northern Kenya, we first document that in age set societies there is strong evidence of within-cohort financial ties and zero evidence of inter-generational investment within the family. In kin-based societies we observe the opposite pattern. Second, analyzing the introduction of Uganda’s national pension program, we show that pension grants had large, positive spillover effects on child nutrition in kin-based societies, but no effect in age set societies. Local social structure leads to strikingly different aggregate and distributional consequences of public policy. Insight from anthropology about variation in social structure and how it shapes the formation of financial ties could contribute in important ways to economic and policy analysis.

36 References

Akbari, Mahsa, Duman Bahrami-Rad, and Erik Kimbrough, “Kinship, Fractionalization and Cor- ruption,” 2017. Working paper.

Alesina, Alberto and Paola Giuliano, “The Power of the Family,” Journal of Economic Growth, June 2010, 15 (2), 93–125.

and , “Family ties,” in “Handbook of economic growth,” Vol. 2, Elsevier, 2014, pp. 177–215.

Ambrus, Attila and Matthew Elliott, “Investments in social ties, risk sharing and inequality,” Eco- nomic Research Initiatives at Duke (ERID) Working Paper, 2020, (179).

, Markus Mobius, and Adam Szeidl, “Consumption Risk-Sharing in Social Networks,” The Amer- ican economic review, 2014, 104 (1), 149–182.

Angelucci, Manuela and Giacomo De Giorgi, “Indirect effects of an aid program: how do cash transfers affect ineligibles’ consumption?,” American economic review, 2009, 99 (1), 486–508.

, , and Imran Rasul, “Consumption and investment in resource pooling family networks,” The Economic Journal, 2018, 128 (615), 2613–2651.

Ashraf, Nava, Natalie Bau, Nathan Nunn, and Alessandra Voena, “Bride price and female education,” Journal of Political Economy, 2020, 128 (2), 591–641.

Baird, Sarah, Craig McIntosh, and Berk Özler, When the money runs out: do cash transfers have sus- tained effects on human capital accumulation?, The World Bank, 2016.

Bantebya, Grace Kyomuhendo, Florence Kyoheirwe Muhanguzi, and Carol Watson, “Adolescent girls in the balance: Changes and continuity in social norms and practices around marriage and education in Uganda,” Overseas Development Institute: London, UK, 2014.

Barcellos, Silvia Helena, Leandro S Carvalho, and Adriana Lleras-Muney, “Child gender and parental investments in India: Are boys and girls treated differently?,” American Economic Jour- nal: Applied Economics, 2014, 6 (1), 157–89.

Bau, Natalie, “Can Policy Crowd Out Culture?,” 2016. Working paper.

, “Can policy change culture? Government pension plans and traditional kinship practices,” 2019.

Baxter, Paul TW et al., “Boran age-sets and warfare,” Senri Ethnological Studies, 1979, 3, 69–95.

Beckingham, Charles Fraser and George Wynn Brereton Huntingford, Some Records of Ethiopia, 1593-1646: Being Extracts from the History of High Ethiopia Or Abassia, by Manoel de Almeida, Together with Bahrey’s History of the Galla number 107, Hakluyt Society, 1954.

Bernardi, Bernardo, Age Class Systems: Social Institutions and Politics Based on Age., Cambridge Uni- versity Press, 1985.

37 Bertrand, Marianne, Sendhil Mullainathan, and Douglas Miller, “Public policy and extended families: Evidence from pensions in South Africa,” the world bank economic review, 2003, 17 (1), 27–50.

Bloch, Francis, Garance Genicot, and Debraj Ray, “Informal insurance in social networks,” Journal of Economic Theory, 2008, 143 (1), 36–58.

Boas, Franz, “Museums of ethnology and their classiﬁcation,” Science, 1887, 9 (229).

Bohannan, Laura, Return to Laughter: An Anthropological Novel, New York: Anchor Books, 1964.

Bramoullé, Yann and Rachel Kranton, “Risk-sharing networks,” Journal of Economic Behavior & Or- ganization, 2007, 64 (3-4), 275–294.

Breza, Emily, Arun Chandrasekhar, Benjamin Golub, and Aneesha Parvathaneni, “Networks in economic development,” Oxford Review of Economic Policy, 2019, 35 (4), 678–721.

Cahyadi, Nur, Rema Hanna, Benjamin A Olken, Rizal Adi Prima, Elan Satriawan, and Ekki Syam- sulhakim, “Cumulative impacts of conditional cash transfer programs: Experimental evidence from Indonesia,” Technical Report, National Bureau of Economic Research 2018.

Cruz, Cesi, Julien Labonne, and Pablo Querubin, “Politician family networks and electoral outcomes: Evidence from the Philippines,” American Economic Review, 2017, 107 (10), 3006–37.

Duﬂo, Esther, “Grandmothers and granddaughters: old-age pensions and intrahousehold allocation in South Africa,” The World Bank Economic Review, 2003, 17 (1), 1–25.

Dyson-Hudson, Neville, “The Karimojong age system,” Ethnology, 1963, 2 (3), 353–401.

Egger, Dennis, Johannes Haushofer, Edward Miguel, Paul Niehaus, and Michael W Walker, “Gen- eral equilibrium effects of cash transfers: experimental evidence from Kenya,” Technical Report, National Bureau of Economic Research 2019.

Eisenstadt, Shmuel N, “African age groups: a comparative study,” Africa, 1954, 24 (2), 100.

Eldredge, Elizabeth A., Creation of the Zulu Kingdom, 1815-1828: War, Shaka, and the Consolidation of Power, New York: Cambridge University Press, 2014.

Enke, Benjamin, “Kinship Systems, Cooperation, and the Evolution of Moral Systems,” Quarterly Journal of Economics, 2019, 134 (2), 953–1019.

Fafchamps, Marcel, Rural poverty, risk and development, Vol. 144, Edward Elgar Publishing, 2003.

, “Risk sharing between households,” in “Handbook of social economics,” Vol. 1, Elsevier, 2011, pp. 1255–1279.

and Flore Gubert, “The formation of risk sharing networks,” Journal of development Economics, 2007, 83 (2), 326–350.

38 Ferrara, Eliana La, “Descent rules and strategic transfers. Evidence from matrilineal groups in Ghana,” Journal of Development Economics, 2007, 83 (2), 280–301.

Foner, Anne and David Kertzer, “Transitions over the life course: Lessons from age-set societies,” American Journal of Sociology, 1978, 83 (5), 1081–1104.

Green, C, A Mukuria, and D Rubin, “Addressing early marriage in Uganda,” Washington DC, US- AID, Futures Group, Health Policy Initiative, Task Order I, 2009.

Greif, Avner, “Cultural Beliefs and the Organization of Society: A Historical and Theoretical Reﬂec- tion on Collectivist and Individualist Societies,” Journal of Political Economy, 1994, 102 (5), 912–950.

Handley, Carla S., “Notes and Interview Transcripts from Northern Kenya,” 2009. Field Notes.

Heusch, Luc De, Sacriﬁce in Africa: a structuralist approach, Manchester University Press, 1985.

Hinew, Dereje, “History of Oromo social organization: Gadaa grades based roles and responsibilities,” Science, Technology and Arts Research Journal, 2012, 1 (3), 88–96.

Hoddinott, John, Jere R Behrman, John A Maluccio, Paul Melgar, Agnes R Quisumbing, Manuel Ramirez-Zea, Aryeh D Stein, Kathryn M Yount, and Reynaldo Martorell, “Adult consequences of growth failure in early childhood,” The American journal of clinical nutrition, 2013, 98 (5), 1170–1178.

Jayachandran, Seema and Rohini Pande, “Why are Indian children so short? The role of birth order and son preference,” American Economic Review, 2017, 107 (9), 2600–2629.

Kelly, Raymond C, “A Note on Nuer Segmentary Organization,” 1983.

Kinnan, Cynthia and Robert Townsend, “Kinship and ﬁnancial networks, formal ﬁnancial access, and risk reduction,” American Economic Review, 2012, 102 (3), 289–93.

, Krislert Samphantharak, Robert M Townsend, and Diego Vera Cossio, “Insurance and Propa- gation in Village Networks,” NBER Working Paper, 2020, (w28089).

Kocherlakota, Narayana R, “Implications of efﬁcient risk sharing without commitment,” The Review of Economic Studies, 1996, 63 (4), 595–609.

La Ferrara, Eliana, “Descent Rules and Strategic Transfers. Evidence from Matrilineal Groups in Ghana,” Journal of Development Economics, 2007, 83 (2), 280–301.

and Annamaria Milazzo, “Customary Norms, Inheritance and Human Capital. Evidence from a Reform of the Matrilineal System in Ghana,” 2011. Working paper.

LeVine, Robert A and Walter H Sangree, “The diffusion of Age-group organization in East Africa. A controlled comparison,” Africa, 1962, pp. 97–110.

Lowes, Sara, “Kinship Systems, Gender Norms, and Household Bargaining: Evidence from the Ma- trilineal Belt,” 2016. Working paper, Harvard University.

39 Lowie, Robert H., Primitive Society, New York: Liveright, 1921.

Lowie, Robert Harry, Culture & ethnology, DC McMurtrie, 1917.

Mazzocco, Maurizio and Shiv Saini, “Testing efﬁcient risk sharing with heterogeneous risk prefer- ences,” American Economic Review, 2012, 102 (1), 428–68.

McCluskey, Kathleen A, Life-span developmental psychology: Historical and generational effects, Elsevier, 2013.

Mei, Zuguo, Cynthia L Ogden, Katherine M Flegal, and Laurence M Grummer-Strawn, “Compar- ison of the prevalence of shortness, underweight, and overweight among US children aged 0 to 59 months by using the CDC 2000 and the WHO 2006 growth charts,” The Journal of pediatrics, 2008, 153 (5), 622–628.

Milazzo, Annamaria, Son preference, fertility and family structure: Evidence from reproductive behavior among Nigerian women, The World Bank, 2014.

Mobarak, Ahmed Mushﬁq and Mark R Rosenzweig, “Informal risk sharing, index insurance, and risk taking in developing countries,” American Economic Review, 2013, 103 (3), 375–80.

Morduch, Jonathan, “Consumption smoothing across space,” Dercon, Insurance against Poverty. Ox- ford University Press, New York and Oxford, 2004.

Morton, RF, “The structure of east African age-set systems,” Pula: Botswana journal of African studies, 1979, 1 (2), 77–102.

Moscona, Jacob, Nathan Nunn, and James A. Robinson, “Keeping It in the Family: Lineage Or- ganization and the Scope of Trust in Sub-Saharan Africa,” American Economic Review Papers and Proceedings, 2017, 107 (5), 565–571.

, , and James A Robinson, “Segmentary Lineage Organization and Conﬂict in Sub-Saharan Africa,” Econometrica, 2020, 88 (5), 1999–2036.

Munshi, Kaivan and Mark Rosenzweig, “Traditional Institutions Meet the Modern World: Caste, Gender, and Schooling Choice in a Globalizing Economy,” American Economic Review, September 2006, 96 (4), 1225–1252.

Murdock, George Peter, “Ethnographic atlas,” 1967.

Naidu, Suresh, James A Robinson, and Lauren E Young, “Social origins of dictatorships: Elite networks and political transitions in Haiti,” Unpublished Manuscript, 2015.

Philips, SR and RJ Shulman, “Measurement of growth in children,” UpToDate. Waltham, MA, 2018.

Radcliffe-Brown, Alfred R., “Age Organization-Terminology,” Man, 1929, 29 (21).

Ravallion, Martin and Shubham Chaudhuri, “Risk and insurance in village India: Comment,” Econometrica: Journal of the Econometric Society, 1997, pp. 171–184.

40 Rosenzweig, Mark R and Oded Stark, “Consumption smoothing, migration, and marriage: Evi- dence from rural India,” Journal of political Economy, 1989, 97 (4), 905–926.

Salzman, Philip C., “Ideology and Change in Middle Eastern Tribal Societies,” Man, 1978, 13 (4), 618–637.

Smith, M.G., “On Segmentary Lineage Systems,” Journal of the Royal Anthropological Institute of Great Britain and Ireland, 1956, 86 (2), 39–80.

Spencer, Paul, Youth and Experiences of Ageing among Maa: Models of Society Evoked by the Maasai, Samburu, and Chamus of Kenya, Walter de Gruyter GmbH & Co KG, 2014.

Townsend, Robert M, “Risk and insurance in village India,” Econometrica: Journal of the Econometric Society, 1994, pp. 539–591.

, “Consumption insurance: An evaluation of risk-bearing systems in low-income economies,” Jour- nal of Economic perspectives, 1995, 9 (3), 83–102.

Udry, Christopher, “Risk and insurance in a rural credit market: An empirical investigation in northern Nigeria,” The Review of Economic Studies, 1994, 61 (3), 495–526.

41 A1 Coding of the Age-Set Variable

For our sample of ethnic groups in both Uganda and Kenya, we used a broad range of ethnographic work to determine whether or not age sets were the dominant form of social organization. Table A1 lists the ethnic groups in our sample, along with whether or not each is an age set society and the references used to make this determination. We also report the share of households in the DHS belonging to each ethnic group; we were able to verify whether or not age set organization was the dominant form of social organization for 85% of the sample. The majority of the remaining ethnic groups are small—under 0.1% of the DHS sample—and we were not able to ﬁnd sufﬁcient information to determine whether or not age sets are present. Table A2 lists the ethnic groups our sample for the analysis of the Hunger Safety Net Program (HSNP) in Northern Kenya, along with whether or not each is an age set society and the references used to make this determination. In the HSNP data, unlike the DHS, language rather than ethnicity is reported. The only language in the sample that is not immediately straightforward to link to a particular ethnic group is Swahili, which is the home language of 0.18% of the sample. Thus, we exclude individuals who report speaking Swahili at home from the analysis; however, this is a very small part of the sample and for the remaining 99.82% of the same we were able to determine whether the household is a member of an age set society or not.

42 Table A1: Identifying Age Set Societies: Uganda

Share of DHS Ethnic Group Sample Agesets References acholi 5.84 Yes Butt, Audrey. The Nilotes of the Sudan and Uganda: East Central Africa Part IV . Southall, Aidan William. Alur Society: A Study in Processes and Types of Domination . LIT alur 2.95 No Verlag Münster, 2004. McCluskey, Kathleen A. Life-Span Developmental Psychology: Historical and baganda 13.16 No Generational Effects . Elsevier, 2013. De Wolf, J. J. “The Diffusion of Age-Group Organization in East Africa: A Reconsideration.” Africa: Journal of the International African Institute 50, no. 3 (1980): bagisu 5.36 No 305–10. https://doi.org/10.2307/1159121. Taylor, Brian K. The Western Lacustrine Bantu (Nyoro, Toro, Nyankore, Kiga, Haya and bakiga 7.67 No Zinza with Sections on the Amba and Konjo): East Central Africa Part XIII . Routledge, Elam, Yitzchak. “Family and Polity in Ankole: The Hima Household and the Absence of banyankore 9.12 No Age-Sets.” Ethnology 14, no. 2 (1975): 163–71. https://doi.org/10.2307/3773087. Lamony, Stephen Arthur. “Approaching National Reconciliation in Uganda: Perspectives on Applicable Justice Systems.” Ugandan Coalition of the International Criminal Court, 2007; Banyole also described as similar in structure to Samia, who do banyole 1.69 No not have age sets (see Samia). Main social structure and organizationn is the clan. Taylor, Brian K. The Western Lacustrine Bantu (Nyoro, Toro, Nyankore, Kiga, Haya and banyoro 3.29 No Zinza with Sections on the Amba and Konjo): East Central Africa Part XIII . Routledge, basamia 1.5 No Sokolovsky, Jay. The Cultural Context of Aging: Worldwide Perspectives . Praeger, 2009. Ochieng’, William Robert. Historical Studies and Social Change in Western Kenya: Essays basoga 6.27 No in Memory of Professor Gideon S. Were . East African Publishers, 2002. Taylor, Brian K. The Western Lacustrine Bantu (Nyoro, Toro, Nyankore, Kiga, Haya and batoro 2.7 No Zinza with Sections on the Amba and Konjo): East Central Africa Part XIII . Routledge, McCluskey, Kathleen A. Life-Span Developmental Psychology: Historical and iteso 8.85 Yes Generational Effects . Elsevier, 2013. Foner, Anne, and David Kertzer. “Transitions Over the Life Course: Lessons from Age- jie 0.66 Yes Set Societies.” American Journal of Sociology 83, no. 5 (1978): 1081–1104. Southall, Aidan William. Alur Society: A Study in Processes and Types of Domination . LIT jonam 0.26 No Verlag Münster, 2004; Also, ub-division of Alur, which do not have age sets (see Alur) Minority Rights Group. “Uganda.” Accessed November 10, 2020. kakwa 0.37 Yes https://minorityrights.org/country/uganda/. Dyson-Hudson, Neville. “The Karimojong Age System.” Ethnology 2, no. 3 (1963): karimojong 1.53 Yes 353–401. https://doi.org/10.2307/3772867. Weatherby, John M. The Sor Or Tepes of Karamoja (Uganda): Aspects of Their History and Culture . Ediciones Universidad de Salamanca, 2012; Also described as closely kumam 0.87 Yes related socials structure to Karimojong, who have age sets. Adt, S. N. Eisenst. “African Age Groups: A Comparative Study.” Africa: Journal of the lango 7.01 Yes International African Institute 24, no. 2 (1954): 100–113. Southall, Aidan William. Alur Society: A Study in Processes and Types of Domination . LIT lendu 0.04 No Verlag Münster, 2004. Weeks, Sheldon G. “Youth and the Transition to Adult Status: Uganda.” Journal of Youth lugbara 2.91 No and Adolescence 2, no. 3 (September 1, 1973): 259–70. Mtodleton, John. “Notes on the Political Organization of the Madi of Uganda.” African madi 1.08 No Studies 14, no. 1 (1955): 29–36. Adt, S. N. Eisenst. “African Age Groups: A Comparative Study.” Africa: Journal of the pokot 0.36 Yes International African Institute 24, no. 2 (1954): 100–113. sabiny 0.61 Yes Goldschmidt, Walter. Sebei Law . University of California Press, 1967. Foner, Anne, and David Kertzer. “Transitions Over the Life Course: Lessons from Age- so (tepeth) 0.03 Yes Set Societies.” American Journal of Sociology 83, no. 5 (1978): 1081–1104.

43 Table A2: Identifying Age Set Societies: Kenya

Share of HSNP Ethnic Group Sample Agesets Reference Legesse, Asmarom. Gada: Three approaches to the study of African society . New borana 12.26 Yes York: Free Press, 1973. Amborn, Hermann, and Ruth Schubert. "The contemporary significance of what has been. Three approaches to remembering the past: Lineage, gada, and oral tradition." History in Africa 33 (2006): 53-84.; Hamer, John H. "Sidamo burji 1.78 No generational class cycles. A political gerontocracy." Africa 40.1 (1970): 50-70. Tablino, Paolo. The Gabra: camel nomads of northern Kenya . No. 4. Paulines gabra 5.49 Yes Publications Africa, 1999. Beaman, Anne W. The Rendille age-set system in ethnographic context: adaptation and integration in a nomadic society. Diss. 1981; Lewis, Ioan M. "Force and fission in northern Somali lineage structure." American Anthropologist (1961): 94-112.; Lewis, Ioan Myrddin. A Pastoral Democracy: a study of pastoralism and politics among the northern Somali of the Horn of garre 7.31 No Africa. James Currey Publishers, 1999. Spencer, Paul. Nomads in alliance: symbiosis and growth among the Rendille and Samburu of Kenya. Oxford University Press, 2012.; Stewart, Frank Henderson. "Fundamentals of Age-Group Systems." (1977).; Beaman, Anne W. The Rendille age-set system in ethnographic context: adaptation and integration in a nomadic society. Diss. 1981.; Handley, Carla S., “Notes and rendille 8.56 Yes Interview Transcripts from Northern Kenya,” 2009. Field Notes. Spencer, Paul. Nomads in alliance: symbiosis and growth among the Rendille and Samburu of Kenya . Oxford University Press, 2012.; Handley, Carla S., samburu 2.87 Yes “Notes and Interview Transcripts from Northern Kenya,” 2009. Field Notes. Beaman, Anne W. The Rendille age-set system in ethnographic context: adaptation and integration in a nomadic society. Diss. 1981.; Lewis, Ioan M. "Force and fission in northern Somali lineage structure." American Anthropologist (1961): 94-112.; Lewis, Ioan Myrddin. A Pastoral Democracy: a study of pastoralism and politics among the northern Somali of the Horn of somali 35 No Africa. James Currey Publishers, 1999. Gulliver, Philip H. "The Turkana age organization." American Anthropologist 60.5 (1958): 900-922.; Handley, Carla S., “Notes and Interview Transcripts turkana 26.54 Yes from Northern Kenya,” 2009. Field Notes.

44 A2 Supplementary Figures

Percentile Bottom 5% Indicator 30 0

-.05 20 -.0596

13.98 -.1 -.0952 -.0998

10 8.722 8.999

-.15

0 -.2 Baseline HH Controls All Controls Baseline HH Controls All Controls

(a) Double Difference: WFH Percentile (b) Double Difference: WFH Bottom 5% Indicator

Figure A1: Double Difference Estimates. In Figure A1a the dependent variable is child weight-for-height percentile and in Figure A1b it is an indicator if the child falls below the 5th percentile. Each column reports our coefficient estimate of the difference in the effect of being in a pension pilot district on households in non- age set vs. age set societies. The leftmost column includes only our baseline controls: ethnicity fixed effects, age-by-sex fixed effects, interview fixed effects, and a pilot district indicator. I the middle column we add our expanded set of household-level controls and in the right column we add our full set of ethnicity-level controls. 95% confidence intervals are reported.

45 h aevlaeadehiiybtotieteaechr ftemi rvdrfrmmeso oite with societies of members for provider of main the members of across cohort average age the respectively. the from respectively. sets outside age sets deviation but without age expenditure and ethnicity without of and and distribution with village the societies same of display the members A2d for and average A2c cohort Figures age provider’s main the from A2: Figure

Percent Percent 0 5 10 15 20 0 20 40 60 80 Mean=0.000, sd=0.196 Mean=0.004, sd=0.598 c usd ootDvain g Set Age Deviation: Cohort Outside (c) osmto Deviations. Consumption -2 -2 a ootDvain g Set Age Deviation: Cohort (a) Deviation fromethnicgroupmean Deviation fromcohortmean -1 0 0 1 dslytedsrbto fepniuedeviation expenditure of distribution the display A2b and A2a Figures 2 2 46

Percent Percent 0 5 10 15 20 0 20 40 60 80 Mean=-0.000, sd=0.223 Mean=0.013, sd=0.554 -4 d usd ootDvain oAeSet Age No Deviation: Cohort Outside (d) b ootDvain oAeSet Age No Deviation: Cohort (b) -2 Deviation fromethnicgroupmean Deviation fromcohortmean -1 0 0 1 2 2 3 4 tmaue lgo)fo xedtr.Bt rpsaerpre fe bobn o fhousehold of log absorbing after reported are squared. in graphs income while Both household expenditure total of expenditure. of) log societies food (log and of measures of) income y-axis members (log the for measures A4a, separately Figure it In (y-axis), A4b , providers sets. Figure age main (orange) household with the and of (blue) without age the vs. (y-axis) expenditure A4: Figure ( slopes two the between difference the (blue) of without p-value societies the for report of (y-axis) also members eligible We for HSNP sets. separately are age plotted who (orange) is providers with relationship main and The the of non-beneﬁciaries. members of cohort sample of the share the vs. (x-axis) expenditure A3: Figure

Log(1+Tot Exp) 8.7 8.8 8.9 9 20 osmto n ootHN Eligibility. HSNP Cohort and Consumption osmto eitosCnrligfrIncome. for Controlling Deviations Consumption a oa Expenditure Total (a) Share Eligible in Cohort 40 No age-sets 0 .2 .4 .6 .8 1 Age ofthehhhead 6.5 β Age-Sets =-0.461, Age-sets β NoAge-Sets 60 7 .=0.107, P-val=0.000 Log(1 +TotalExpenditures) No Age-Sets 80 47 7.5 Log(1+Food Exp) isatrposo lgo)hueodconsumption household of) (log of plots Binscatter 8.4 8.5 8.6 8.7 20 Age-Sets isatrposo oshl consumption household of plots Binscatter b odExpenditure Food (b) 8 40 No age-sets Age ofthehhhead Age-sets 8.5 60 p <0.01). 80 log Total Spending

1.5

1 .9245

.5416 .5

.0883 0

-.5 Under 30 Age 30-50 Over 50

(a) Total Expenditure

log Food Spending 1.5

1 .7948 .6214 .5

.1551

-.5 Under 30 Age 30-50 Over 50

(b) Food Expenditure

Figure A5: Heterogeneous Effects By Generation. Estimates of the coefﬁcient of interest from Equation2, estimated separately on a sample of main providers under 30 years old, from 30-50 years old, and over 50 years old. In Figure A5a, the y-axis measures (log of) total expenditure while in Figure A5b, it measures (log of) food expenditure. 95% conﬁdence intervals are reported.

48 A3 Supplementary Tables

Table A3: Age Cohort Spillover Effects On Expenditure: Cohort of the Household Heads and Main Providers

Dep var: Log (1+ Per Capita Total Expenditure) (1) (2) (3) (4) Full Sample Males Only Full Sample Males Only Household Head Main Provider & Hh Head

Share of elig in Age-Set * ITreat * IAge−Set (v) 0.246** 0.329*** 0.266** 0.344*** (0.110) (0.112) (0.119) (0.121) Share of elig in Age-Set * ITreat * INoAge−Set (v) 0.0423 0.0287 0.0538 0.0277 (0.0861) (0.0891) (0.0925) (0.0945)

Observations 685 577 621 546 R-squared 0.599 0.627 0.616 0.626 Interviewer FE Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Age FE Yes Yes Yes Yes Age-set X Share of Age-Set Eligible Yes Yes Yes Yes Mean at baseline 7.317 7.344 7.337 7.356 P-val γ1 = γ2 0.180 0.0492 0.200 0.0539 Notes: Share of elig in Age-Set is the share of eligible age-mates of the household head. Log per capita monthly expenditures. IAge−Set is a variable that takes value one if the respondent belongs to an ethnicity that has age-sets and 0 otherwise, INoAge−Set is a variable that takes value one if the person belongs to an ethnicity that does not have age-sets. Finally, ITreat is a dummy that takes value one if the respondent is in the treatment group and 0 otherwise, We control for economic and demographic variables such as gender, education, marriage status, occupation, religion, household size, size of the cohort, and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

49 Table A4: Age Cohort Spillover Effects On Food Expenditure

Dep Var: Log(1+Per Capita Food Expenditure) (1) (2) (3) (4) (5) (6) Full Sample Males Only

Share of elig in Age-Set * ITreat * IAge−Set 0.205** 0.261*** 0.276*** 0.276** 0.299*** 0.339*** (0.0919) (0.0950) (0.0977) (0.108) (0.106) (0.108) Share of elig in Age-Set * ITreat * INoAge−Set -0.0196 -0.0349 -0.0327 -0.00176 -0.0317 -0.0338 (0.0698) (0.0879) (0.0893) (0.0678) (0.0861) (0.0893) p-value, coefficient difference 0.0630 0.0366 0.0308 0.0391 0.0249 0.0105 Observations 713 707 704 603 596 593 R-squared 0.473 0.519 0.527 0.473 0.531 0.544 Interviewer FE Yes Yes Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Yes Yes Baseline Controls Yes Yes Yes Yes Yes Yes Age FE No Yes Yes No Yes Yes Age-set X Share of Age-Set Eligible No No Yes No No Yes Mean at baseline 7.041 7.040 7.038 7.353 7.354 7.352 Notes: The unit of observation is a non-eligible household. Baseline controls include gender, education, marriage status, occupation, religion, household size and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

50 Table A5: Effect of Age-Set Treatment on Cash Transfers among households

(1) (2) (3) (4) (5) (6) Received transfer Total received (KSh)

Share of elig in Age-Set * ITreat * IAge−Set -0.00507 0.0117 0.0288 559.7 763.0* 841.8* (0.0824) (0.106) (0.113) (389.5) (419.4) (441.4) Share of elig in Age-Set * ITreat * INoAge−Set 0.0972 0.0542 0.0569 255.2 160.8 235.6 (0.128) (0.164) (0.176) (337.9) (371.8) (397.4) Share of elig in Age-Set * IAge−Set 0.173** 0.219*** -414.0 -333.9 (0.0694) (0.0736) (272.9) (256.1) Share of elig in Age-Set * INoAge−Set 0.0478 0.0869 314.5 406.4 (0.106) (0.134) (246.7) (295.4)

Observations 570 560 557 570 560 557 R-squared 0.377 0.463 0.469 0.294 0.375 0.379 Interviewer FE Yes Yes Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Yes Yes Age FE No Yes Yes No Yes Yes Age-set X Share of Age-Set Eligible No No Yes No No Yes Mean at baseline 0.414 0.411 0.411 2032 2033 2043 P-val γ1 = γ2 0.484 0.817 0.889 0.548 0.297 0.334 Notes: The table reports the effect of age-sets on transfers executed over the past 3 months. The sample is comprised of all male main respondents. We look at the effects on two dependent variable. Received transfers is a dummy that takes value one if the respondent receives cash or in-kind transfers from another household and 0 otherwise. Column 2 and 4 instead report the effects on the total value received, in-kind or cash, in Kenyan Shillings. Additionally, IAge−Set is a variable that takes value one if the respondent belongs to an ethnicity that has age-sets and 0 otherwise, INoAge−Set is a variable that takes value one if the person belongs to an ethnicity that does not have age-sets. Finally, T is a dummy that takes value one if the respondent is in the treatment group and 0 otherwise, We control for economic and demographic variables such as gender, education, marriage status, occupation, religion, household size, size of the cohort, and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

51 Table A6: Controlling for Proxies for the Family

Dep Var: Log(1+Total Per Capita Expenditures) (1) (2) (3) (4) Full Sample Males Only Full Sample Males Only Ethnicity Clan

Share of elig in Age-Set * ITreat* IAge−Set 0.308** 0.429*** 0.455*** 0.551*** (0.116) (0.115) (0.138) (0.153) Share of elig in Age-Set * ITreat* INoAge−Set 0.0412 -0.00335 0.0328 0.0111 (0.0771) (0.0838) (0.100) (0.105) Share of elig in Ethn * ITreat * IAge−Set -0.329 -0.573* (0.287) (0.331) Share of elig in Ethn * ITreat * INoAge−Set -0.320 -0.0802 (0.255) (0.246) Share of elig in Clan * ITreat* IAge−Set -0.514** -0.652** (0.255) (0.261) Share of elig in Clan * ITreat * INoAge−Set -0.169 -0.0323 (0.231) (0.205)

Observations 704 593 556 477 R-squared 0.603 0.621 0.615 0.632 Interviewer FE Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Age FE Yes Yes Yes Yes Age-set X Share of Age-Set Eligible Yes Yes Yes Yes Mean at baseline 7.318 7.352 7.323 7.346 P-val γ1 = γ2 0.0826 0.00765 0.0209 0.00812 P-val φ1 = φ2 0.980 0.231 0.314 0.0580 P-val γ1 + φ1 = 0 0.929 0.625 0.764 0.626 P-val γ2 + φ2 = 0 0.256 0.705 0.531 0.911 IAge−Set is a variable that takes value one if the respondent belongs to an ethnicity that has age-sets and 0 otherwise, INoAge−Set is a variable that takes value one if the person belongs to an ethnicity that does not have age-sets. Finally, ITreat is a dummy that takes value one if the respondent is in the treatment group and 0 otherwise, We control for economic and demographic variables such as gender, education, marriage status, occupation, religion, household size and a poverty index, size of the age cohort and size of the clan/ethnic group. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. We denote by γ1 and γ2 the coefficients on the first two rows, φ1 and φ2, the coefficients on the second two clan/ ethnic group. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

52 Table A7: Robustness to Controlling for Ethnicity Level Characteristics

Dep Var: Log(1+Total per-capita expenditures) (1) (2) (3) (4) (5) (6) (7) (8) Full Sample Males Only

Share of elig in Age-Set * ITreat * IAge−Set 0.286*** 0.258** 1.132*** 1.130*** 0.387*** 0.341*** 1.228*** 1.225*** (0.0994) (0.0997) (0.204) (0.204) (0.107) (0.104) (0.222) (0.222) Share of elig in Age-Set * ITreat * INoAge−Set -0.0411 -0.0146 0.822*** 0.820*** -0.0513 -0.0332 0.872*** 0.875*** (0.0924) (0.0885) (0.175) (0.177) (0.0856) (0.0827) (0.174) (0.175)

Observations 704 704 704 704 593 593 593 593 R-squared 0.539 0.566 0.576 0.576 0.573 0.592 0.600 0.600 53 Interviewer FE Yes Yes Yes Yes Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Yes Yes Yes Yes Age FE Yes Yes Yes Yes Yes Yes Yes Yes Age-set X Share of Age-Set Eligible Yes Yes Yes Yes Yes Yes Yes Yes Houeshold level Controls No Yes No No No Yes No No Ethnicity level controls No No Yes Yes No No Yes Yes Language family level controls No No No Yes No No No Yes Mean at baseline 7.318 7.318 7.318 7.318 7.352 7.352 7.352 7.352 P-val γ1 = γ2 0.0276 0.0657 0.0460 0.0453 0.00284 0.0104 0.0225 0.0294 Notes: In the first four columns, the sample includes all the household included in the evaluation, in columns 5-8 the sample includes only the males main respondents. The dependent variable is Log of total per capita annual expenditures. The estimated specification is Equation2 to which we add several controls for robustness. Columns 1 and 4 report the estimated coefficients, with no controls besides the lagged value of the dependent value and the preferred fixed effects. Columns 2 and 4 include household level controls to the equation, this is our preferred specification. Columns 3 and 5, includes ethnicity level characteristics as measured in the Murdock Ethnographic Atlas. Finally, columns 4 and 8 includes controls at the language-family level. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1. Table A8: Life Cycle Consumption in Societies With and Without Age Sets

Dep Var: Log(1+Total Expenditures) (1) (2) (3) (4) Full Sample Males Only Full Sample Males Only Main Providers Main providers who are hh heads

Age*IAge−Set 0.00534 0.00636 0.00925** 0.0127*** (0.00372) (0.00451) (0.00447) (0.00468) Age*INoAge−Set -0.0121*** -0.0131*** -0.0144*** -0.0135** (0.00396) (0.00428) (0.00497) (0.00508) Age2*IAge−Set -7.32e-05* -7.09e-05 -0.000119** -0.000139*** (4.07e-05) (4.87e-05) (4.75e-05) (4.90e-05) Age2*INoAge−Set 9.95e-05** 0.000118*** 0.000118** 0.000119** (4.16e-05) (4.11e-05) (5.04e-05) (4.80e-05)

Observations 5,063 3,812 4,224 3,326 R-squared 0.519 0.509 0.517 0.502 Sub-Location x Ethnicity FE Yes Yes Yes Yes Controlling for hh size Yes Yes Yes Yes P-value equality linear Term 0.00215 0.00326 0.000708 0.000442 P-value equality quadratic term 0.00481 0.00531 0.00127 0.000528 Notes: IAge−Set is a variable that takes value one if the respondent belongs to an ethnicity that has age- sets and 0 otherwise, INoAge−Set is a variable that takes value one if the person belongs to an ethnicity that does not have age-sets. The unit of observation is an individual. We estimate the regression for main providers only, and for main providers who are also heads of household. For the latter, age should be even more important (since both main provider and household head belong to the same age-set, by deﬁnition). We control for household size at baseline. Standard errors clustered at the sub-location level in parentheses. *** p<0.01, ** p<0.05, * p<0.1

54 Table A9: Main results excluding CBT sub-locations

Dep Var: Log(1+Total Expenditure) (1) (2) (3) (4) Full Sample Males Only All No CBT All No CBT

Share of elig in Age-Set * T * IAge−Set (v) 0.274*** 0.326** 0.339*** 0.403*** (0.100) (0.128) (0.107) (0.135) Share of elig in Age-Set * T * INoAge−Set (v) -0.0110 0.0681 -0.0191 0.0106 (0.0824) (0.104) (0.0787) (0.114)

Observations 704 476 593 393 R-squared 0.598 0.627 0.616 0.644 Interviewer FE Yes Yes Yes Yes TargetingCode X Ethnicity FE Yes Yes Yes Yes Age FE Yes Yes Yes Yes Age-set X Share of Age-Set Eligible Yes Yes Yes Yes Mean at baseline 7.318 7.289 7.352 7.315 P-val γ1 = γ2 0.0445 0.146 0.0118 0.0363 Notes: Share of elig in Age-Set is the share of eligible age-mates of the household head. Log per capita monthly expenditures. IAge−Set is a variable that takes value one if the respondent belongs to an ethnicity that has age-sets and 0 otherwise, INoAge−Set is a variable that takes value one if the person belongs to an ethnicity that does not have age-sets. Finally, ITreat is a dummy that takes value one if the respondent is in the treatment group and 0 otherwise. The sample excludes sub-locations in which the targeting mechanism was Community Based Targeting. We control for economic and demographic variables such as gender, education, marriage status, occupation, religion, household size, size of the cohort, and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

55 Table A10: Pension Receipt and Child Weight-for-Height: Kenya’s HSNP

Dep Var: Child Weight-For-Height (1) (2) (3) (4) (5) (6) Full Sample Under 5 Male Children Only Percentile Z-Score Bottom 5% Percentile Z-Score Bottom 5% Indicator Indicator ITreat ∗ INoAge−Set 7.874*** 0.394** 0.00529 12.32** 0.478* 0.137 (2.386) (0.164) (0.0553) (4.199) (0.261) (0.134) ITreat ∗ IAge−Set 1.219 0.0594 -0.0726 -3.810 -0.145 -0.0259 (2.658) (0.178) (0.0450) (4.998) (0.251) (0.0558)

Interviewer Fixed Effects Yes Yes Yes Yes Yes Yes Age x Age Set Yes Yes Yes Yes Yes Yes Basline Controls Yes Yes Yes Yes Yes Yes Observations 231 231 231 116 116 116 R-squared 0.359 0.360 0.212 0.440 0.475 0.476 Mean at baseline 50.46 -1.228 0.0390 51.92 -1.193 0.0431 p-value, coefficient difference 0.0634 0.162 0.293 0.0395 0.0997 0.292 Notes: The unit of observation is a child under 5 in the pension targeting mechanism group. Base- line controls include gender, education, marriage status, occupation, religion, household size and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

56 Table A11: Pension Receipt and Education Spending: Kenya’s HSNP

(1) (2) (3) (4) (5) Log(1+Ed. Log(Ed Log(Food Log(Rent Log(Health Spending Spend/ Spend/ Spend/ Spend/ Tot. Spend) Tot. Spend) Tot. Spend) Tot. Spend)

Panel A: Full Sample

ITreat ∗ INoAge−Set 0.990** 0.968** 0.00414 -0.171* 0.500 (0.395) (0.346) (0.0186) (0.0869) (0.404) ITreat ∗ IAge−Set 0.135 0.0591 0.0152 -0.143 -0.148 (0.447) (0.414) (0.0221) (0.169) (0.338) R-squared 0.587 0.575 0.619 0.514 0.409 p-value, coefﬁcient difference 0.139 0.0784 0.655 0.901 0.210

Panel B: Males Only

ITreat ∗ INoAge−Set 1.091* 1.099* 0.00393 -0.108 0.436 (0.588) (0.522) (0.0172) (0.0856) (0.441) ITreat ∗ IAge−Set -0.497 -0.490 0.0156 -0.124 -0.527 (0.629) (0.569) (0.0178) (0.194) (0.360) R-squared 0.584 0.575 0.606 0.556 0.429 p-value, coefﬁcient difference 0.0525 0.0292 0.601 0.945 0.114

Interviewer Fixed Effects Yes Yes Yes Yes Yes Ethnicity x Targeting Code Fixed Effects Yes Yes Yes Yes Yes Age x Age Set Yes Yes Yes Yes Yes Basline Controls Yes Yes Yes Yes Yes Notes: The unit of observation is a household in the pension targeting mechanism group. Baseline controls include gender, education, marriage status, occupation, religion, household size and a poverty index. We also include ethnicity fixed effect, ethnicity x targeting code fixed effects, targeting code fixed effects and age fixed effects. Standard errors are clustered at the sub-location level. *** p<0.01, ** p<0.05, * p<0.1.

57 Table A12: Balance: Societies With vs. Without Age Sets in the DHS

(1) (2) (3) (4) (5) (6) Variable Name Sample Age Set vs. Variable Name Sample Age Set Mean No Age Set Mean No Age Set

Panel A: Household-Level Variables

Pension Exposure (all years) 0.333 0.0985 Pension Exposure (2016) 0.0810 0.0273 (0.147) (0.0289) Pension Exposure (2015) 0.0719 0.0244 Pension Exposure (2014) 0.0666 0.0247 (0.0279) (0.0265) Pension Exposure (2013) 0.0620 0.0116 Pension Exposure (2012) 0.0546 0.0108 (0.0325) (0.0322) Pension Exposure (2011) 0.0513 -0.000118 Children Under 5 1.681 0.0632 (0.0304) (0.0510) Pension-Years Per Child 0.223 0.0276 (0.0861)

Panel B: Child-Level Variables

Weight-for-Height (Percentile) 47.27 0.532 Weight-for-Height (Bottom 5%) 0.0551 -0.0169 (2.212) (0.0182) Weight-for-Age (Percentile) 31.54 1.402 Weight-for-Age (Bottom 5%) 0.208 -0.0392 (3.687) (0.0357) Height-for-Age (Percentile) 27.03 2.300 Height-for-Age (Bottom 5%) 0.315 0.000703 (4.017) (0.0444) Age (Years) 2.453 0.142 (0.0910) Notes: The unit of observation is an household in Panel A and a child in Panel B. The sample is restricted to households and children in non-pilot districts. Columns 1 and 4 report the household of child-level characteristics; columns 2 and 5 report the sample mean of each measure; and columns 3 and 6 report the coefficient on the age set indicator. District and interview month fixed effects are included in all specifications. Standard errors, clustered by ethnicity, are reported in parentheses. *** p<0.01, ** p<0.05, * p<0.1

58 Table A13: Effects of Pension Exposure on Nutrition: Alternative Measures of Nutrition

(1) (2) (3) (4) (5) (6) (7) (8) (9) (10) Dep Var: Weight-for-Height: Weight-for-Age: Height-for-Age: PCA

Percentile Z-Score Bottom 5% Percentile Z-Score Bottom 5% Percentile Z-Score Bottom 5% First PC Indicator Indicator Indicator

Potential Exposure * IPilot* INoAge−Set 3.135*** 0.115** -0.0276*** 2.091* 0.0924** -0.0400*** 0.751 0.0177 -0.0205 0.171*** (1.098) (0.0444) (0.00884) (1.235) (0.0363) (0.0115) (1.213) (0.0282) (0.0126) (0.0617) IPilot I•Age−Set 59 Potential Exposure * * -0.933 -0.0421 0.0302 -1.083 -0.0154 0.0278 0.272 0.0383 -0.0144 -0.0609 (0.951) (0.0390) (0.0199) (0.784) (0.0326) (0.0195) (1.105) (0.0515) (0.0189) (0.0654)

District x Ethnicity Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Interview Month Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Age Set x Potential Exposure Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Age in Months x Gender Fixed Effects Yes Yes Yes Yes Yes Yes Yes Yes Yes Yes Observations 4,107 4,050 4,107 4,068 4,012 4,068 4,068 4,012 4,068 4,012 R-squared 0.203 0.194 0.134 0.256 0.239 0.165 0.229 0.221 0.191 0.240 p-value, coefﬁcient difference 0.00659 0.00958 0.0105 0.0322 0.0338 0.00380 0.767 0.726 0.787 0.0137 Notes: The sample includes all children in the household who are less than 60 months of age. The dependent variable from each regression is noted at the top of each column. In column 10, the dependent variable is the ﬁrst principal component from a principal components analysis including the dependent variables from columns 1-9. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1 Table A14: Effects of Pension Exposure on Nutrition: Alternative Measures of Exposure

Dep Var: Child Weight-for-Height (percentile) (1) (2) (3) (4) Measure of Pension Exposure: Only 2016 Incorporating Child Age

Potential Exposure * IPilot* INoAge−Set 6.247** 10.83** 1.906** 2.481** (2.432) (4.505) (0.844) (1.095) Potential Exposure *IPilot* IAge−Set -1.423 -2.505 -0.694 -0.706 (1.862) (1.986) (0.725) (0.781) p-value, coefﬁcient difference 0.0126 0.00800 0.0197 0.0209 District x Ethnicity Fixed Effects Yes Yes Yes Yes Interview Month Fixed Effects Yes Yes Yes Yes Age Set x Potential Exposure Fixed Effects No Yes No Yes Age in Months x Gender Fixed Effects No Yes No Yes Observations 4,112 4,107 4,112 4,107 R-squared 0.129 0.202 0.129 0.202 Notes: The sample includes all children in the household who are less than 60 months of age. The dependent variable is the child’s weight-for-height percentile. In columns 1-2, pension exposure is computed using only household composition in 2016, and in columns 3-4 it is computed incorporating child age. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1

Table A15: Factor Loadings from Principal Component Analysis

Loading of Variable First Principal Componend Weight-for-Height (percentile) 0.3017 Weight-for-Age (percentile) 0.4233 Height-for-Age (percentile) 0.3124 Weight-for-Height (z-score) 0.3015 Weight-for-Age (z-score) 0.4469 Height-for-Age (z-score) 0.338 Weight-for-Height (Bottom 5% Indicator) -0.1559 Weight-for-Age (Bottom 5% Indicator) -0.3453 Height-for-Age (Bottom 5% Indicator) -0.2895 Notes: This table presents the loading weights of the ﬁrst principal component of the nine characteristics listed in the left column. The sample includes all children under the age of 5 in the 2016 round of Uganda’s DHS.

60 Table A16: Effects of Pension Program on Child Nutrition: Alternative Mechanisms and Additional Controls

(1) (2) (3) (4) (5) (6) (7) Dependent Variable is Child Weight-for-Height (percentile)

Potential Exposure * IPilot* INoAge−Set 3.571*** 3.505*** 3.100*** 4.537** 4.832*** 3.334*** 5.382** (1.269) (1.207) (1.103) (1.755) (1.836) (1.096) (2.232) Potential Exposure *IPilot* IAge−Set -0.881 -0.888 -0.866 -1.347 -2.432 -0.713 -1.956 (0.968) (1.009) (0.934) (2.012) (2.045) (0.908) (1.654)

p-value, coefficient difference 0.00611 0.00709 0.00831 0.0298 0.0111 0.00587 0.00870 Number of Children Fixed Effects Yes No No No No No Yes Household Asset Controls No Yes No No No No Yes Household Wealth Index Fixed Effects No No Yes No No No Yes Ethnicity-Level Controls No No No Yes No No Yes 61 Language-Family Controls No No No No Yes No Yes Interviewer Fixed Effects No No No No No Yes Yes District x Ethnicity Fixed Effects Yes Yes Yes Yes Yes Yes Yes Interview Month Fixed Effects Yes Yes Yes Yes Yes Yes Yes Age Set x Potential Exposure Fixed Effects Yes Yes Yes Yes Yes Yes Yes Age in Months x Gender Fixed Effects Yes Yes Yes Yes Yes Yes Yes Observations 4,107 4,104 4,107 4,107 4,107 4,100 4,097 R-squared 0.206 0.219 0.205 0.204 0.204 0.224 0.246 p-value 0.00611 0.00709 0.00831 0.0298 0.0111 0.00587 0.00870 Notes: The sample includes all children in the household who are less than 60 months of age. The dependent variable is the child’s weight-for-height percentile. Household asset controls include indicators for the presence of electricity, radio, television, a refrigerator, a bicycle, a motorcycle, and a car or truck, as well as fixed effects for main floor material, main roof material, main wall material, the type of toilet facility, and the number of rooms used for sleeping. The household wealth index is an index compuited by the DHS that ranges from 1-5. Ethnicity level controls include fixeed effects for the number of levels of jurisdicrional hierarchy beyond the local community and the ethnic group’s settlement pattern complexity, along with their full set of interactions. Language family fixed effects include each gruop’s language family from the Ethnographic Atlas (v99), along with their full set of interactions. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1 Table A17: Effects of Pension Program on Primary Education in Societies With and Without Age Sets

Dep Var= Currently attending school (1) (2) (3) (4) (5) Full Sample Male Female

Potential Exposure * IPilot* INoAge−Set 0.0104** 0.0102** 0.00661 0.0108** -0.00500 (0.00472) (0.00449) (0.00444) (0.00415) (0.00679) Potential Exposure * IPilot* IAge−Set 0.000619 -0.00317 -0.00347 -0.0190 0.00679 (0.0124) (0.0164) (0.0160) (0.0125) (0.0194)

Observations 18,384 18,383 18,383 9,016 9,223

62 R-squared 0.190 0.191 0.232 0.212 0.271 District x Ethnicity FE Yes Yes Yes Yes Yes Interview Month FE Yes Yes Yes Yes Yes Age Set x Potential Exposure FE No Yes Yes Yes Yes Age x Gender FE No No Yes Yes Yes P-val β = γ 0.464 0.435 0.544 0.026 0.566 Mean Non-Pilot 0.927 0.927 0.927 0.927 0.927 Note: The table reports the estimated coefﬁcients from equation5. The unit of observation is the child. The sample includes all children in the household who are between 6 and 14 years old. The dependent variable is a variable that takes value one if the child is currently in school. Potential exposure is a measure of exposure to the pension constructed as indicated in equation6. ITreat is a dummy that takes value one if the household is in a pilot district, and 0 otherwise. IAge−Set is a variable that takes value one if the household belongs to an ethnicity that has age-sets and 0 otherwise. IAge−Set is a dummy that takes value one if the household belongs to an ethnicity that doesn’t have age-sets and 0 otherwise. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1 Table A18: Effects of Pension Program on Secondary Education and Marriage in Societies With and Without Age Sets

Males Females (1) (2) (3) (4) In school Married In school Married

Potential Exposure * IPilot* INoAge−Set 0.0416* -0.000606 -0.0558*** 0.0381*** (0.0242) (0.00138) (0.0213) (0.0126) Potential Exposure * IPilot* IAge−Set 0.0130 -0.00327 0.00344 0.00313 (0.0302) (0.00343) (0.0186) (0.00886)

Observations 1,740 1,740 2,245 2,245 R-squared 0.205 0.110 0.244 0.190 63 District x Ethnicity FE Yes Yes Yes Yes Interview Month FE Yes Yes Yes Yes Age Set x Potential Exposure FE Yes Yes Yes Yes Age x Gender FE Yes Yes Yes Yes P-val β = γ 0.477 0.475 0.038 0.025 Mean Non-Pilot 0.716 0.040 0.716 0.040 Note: The table reports the estimated coefﬁcients from equation5. The unit of observation is the child. The sample includes all children in the household who are between 15 and 18 years old. The dependent variable are a dummy that takes value one if the child is currently in school and a dummy that takes value 1 if the child is married. Potential exposure is a measure of exposure to the pension constructed as indicated in equation6. ITreat is a dummy that takes value one if the household is in a pilot district, and 0 otherwise. IAge−Set is a variable that takes value one if the household belongs to an ethnicity that has age-sets and 0 otherwise. IAge−Set is a dummy that takes value one if the household belongs to an ethnicity that doesn’t have age-sets and 0 otherwise. Standard errors are clustered at the district level. *** p<0.01, ** p<0.05, * p<0.1