WPS8176 Policy Research Working Paper 8176 Will Elders Provide for Their Grandchildren? Unconditional Cash Transfers and Educational Expenditures in Bolivia Gustavo Canavire-Bacarreza Alberto Chong Fernando Rios-Avila Monica Yanez-Pagans Education Global Practice Group August 2017 Policy Research Working Paper 8176 Abstract This paper takes advantage of repeated cross-section house- the program at the aggregate level. It also finds that the hold surveys and a sharp discontinuity created by the program has stronger effects in indigenous populations introduction of an unconditional cash transfer to elders. and among female and rural populations. The results are The paper evaluates the impact of these cash transfers on robust with respect to a series of falsification tests, survey the educational expenditures for children within a house- structures, model specifications, and estimation methods. hold. The analysis finds positive and significant effects of This paper is a product of the Education Global Practice Group. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://econ.worldbank.org. The authors may be contacted at myanezpagans@ worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Will Elders Provide for Their Grandchildren? Unconditional Cash Transfers and Educational Expenditures in Bolivia Gustavo Canavire-Bacarreza∗ Alberto Chong† Fernando Ríos-Avila ‡ Monica Yanez-Pagans § JEL Classi cation Codes: H55, O15, I12, D12. Keywords: unconditional cash transfers, education, expenditures, children, household, Bolivia. ∗ Universidad EAFIT, Medellín, Colombia. † Georgia State University, Atlanta - USA and Universidad del Pací co, Lima – Peru ‡ Levy Economics Institute, New York – USA § The World Bank, Washington, DC - USA The standard disclaimer applies. 1 Introduction Even though conditional cash transfers (CCTs) have been quite successful in helping achieve devel- opment goals, there has been recent debate on the merits of unconditional cash transfers (UCTs) in relation to conditional ones. It is unclear whether CCTs are a strongly dominating strategy, as UCTs may provide equal or even superior results in some cases. In fact, UCTs have been found to reduce child labor, increase schooling, and improve health and nutrition (Baird et al., 2011). Given that from a policy perspective they are administratively easier to implement and are less costly than CCTs, a clear identi cation of the key characteristics associated with the successful implementation of UCTs is relevant. There are some indications along these lines that are provided by Burlando (2014) and Agüero et al. (2005), who found that attaching conditions to transfers may be super uous providing that a clear alignment of incentives occurs. In this paper, we study the causal impact of UCTs on intra-household allocation in a context in which, given an alignment of incentives, conditions may not be needed. The speci c UCT that we ex- plore arises from an exogenous policy change implemented by the government of Bolivia, by which elderly adults are to be provided with a permanent unconditional transfer of cash. Similar to most developing countries, elderly adults are among the most vulnerable groups in Bolivia, typically due to liquidity constraints that surge from inadequate pension systems and high levels of informality. For the purposes of our research, we focus on the case of extended households where children and grandparents live together in the same residence, something not unusual in many developing coun- tries. It is reasonable to expect that elders living with their grandchildren in the same household will develop stronger bonds and become more invested in the welfare of the children, particularly with respect to human capital investments, such as education and health, which reinforces an alignment of incentives among members of the household. As in most overlapping generation models, an intergenerational transmission of wealth from older generations to younger generations is expected to occur given the alignment of incentives of the older generations with respect to the expectations of welfare improvement of younger generations. In the context of a standard rational expectations approach, this may translate to an increased incentive 2 for the intra-household reallocation of funds from the older generation to the younger in the short term, even when the older generation is facing budget constraints (Barro, 2004; Azariadis, 1993). Interestingly, this idea has rarely been tested. Will the older generations provide for the younger ones in the short term when they are still alive and under constraints? If so, will they do so unconditionally, given the alignment of incentives? There is limited empirical evidence for the e ects of old-age cash transfers and even less on the impact of old-age cash transfers on intra-household investment allocation. As mentioned, the e ect of old-age transfers on intra-household investment allocation is of particular interest given the increasing use of these transfers in developing countries, the number of bene ciaries that live within family households and the potential dependence on the transfers in poorer households. Some related studies include Du o (2003) and Edmonds (2006), who examine the South African old-age pension program and nd evidence that the redistributive consequences of the transfer of food, clothing, and housing are better conditions for households with children. In Latin America, there is some evidence that transfers have reduced poverty (Joubert and Todd, 2011; Barrientos, 2003) as well as evidence regarding intra-household behaviors among households in Chile and Mexico that receiving transfers deviates expenditures towards human capital investments, particularly health (Amuedo-Dorantes and Juarez, 2015; Behrman, 2011). The implementation of an old-age UCT program in Bolivia (Bolivida) provides an excellent oppor- tunity to study the causal e ects of UCTs on intra-household income allocation towards children’s human capital investments, particularly that of education, for which there is, along with health, limited evidence. This program, rst implemented at the end of 2000, has been the core of old-age population support strategies in the country. However, its impacts, if any, are still grossly understud- ied. We take advantage of the fact that the probability of receiving this transfer changes discontinu- ously at the age of eligibility and use this sharp discontinuity to parametrically identify conditional average intent-to-treat e ects and study intra-household income allocation patterns with emphasis on education investments. We measure the impact of this program through child-level educational expenditures, along with old-age program eligibility, after controlling for socioeconomic and demo- 3 graphic characteristics. The eligibility e ect, which we allow to vary across households and children, is assumed to be a log linear function of the ethnicity and gender of the potential recipient. In ad- dition, given the high diversity in the country and the predominance of indigenous populations, we also estimate a series of heterogeneous estimates. Our main nding is that unconditional cash trans- fers to elders lead to substantial improvements in children’s educational investments. We nd that unconditional transfers increased such investments by approximately 60 percent. With respect to ethnicity, eligibility leads to greater increases in schooling investments among indigenous house- holds compared to their non-indigenous counterparts. We nd a similar result for rural and female students.1 Our paper is organized as follows. Section 2 brie y describes the old-age unconditional cash transfer program and its basic characteristics. Section 3 presents the data and basic statistics. Section 4 describes the methodological approach, and Section 5 presents our results. Section 6 then presents robustness tests, and Section 7 concludes the paper. 2 Brief Description of UCT Program for Elders Bolivia is the poorest country in South America, as illustrated by the fact that approximately two- thirds of the households are below the poverty line (The World Bank, 2016). Traditionally, the elderly and the youngest have been the most vulnerable and unprotected segments of the population. In fact, whereas 35 percent of adults live in poverty, nearly 55 percent of children and 60 percent of the elderly do so (The World Bank, 2016), and it is this latter group that is the most unprotected and vulnerable. Moreover, coverage in Bolivia is one of the lowest in the region, with approximately 80 percent of the population having no access to any type of pension system (Landa and Yanez-Pagans, 2008). In addition, the country has a rather complex multiethnic dimension re ected in the fact that it has the 1 Only a handful of studies have examined e ects of this unconditional cash transfer. They have focused on impacts on poverty (Escobar-Loza et al., 2013), household expenditures (Martinez, 2004), co-habitation (Valencia, 2011), health, labor and unintended impacts (Hernani-Limarino and Mena, 2015). The latter provides some basic evidence on education, but exploits a di erent exogenous shock for di erent years. In particular, they take advantage of a reduction in the age of eligibility from 65 to 60 years in 2007. 4 highest percentage of indigenous populations in Latin America, with almost 50 percent of Bolivians self-reporting as belonging to an indigenous group.2 As part of the structural reforms of the 1990s that were broadly supported by the World Bank and the International Monetary Fund, the government created an old-age unconditional cash trans- fer program that entitled all Bolivians aged 65 and older to receive a at, noncontributory, uncondi- tional cash transfer independent of their income levels. In short, the only eligibility rule was age3 . The government determined that this pension would be nanced with the dividends of the shares of privatized companies as part of the structural reform e ort. Whereas the original program, known as Bonosol (Bono Solidario), was rst introduced in 1997, liquidity problems were so serious that in 1998, it had to be put on hold, and when it resumed, at the end of 2000, it was renamed Bolivida, and the pension was reduced from annuities of US$248 to US$60. Figure 1 presents estimates by bene- ciaries. Though take-up is high at the margin of eligibility, compliance is not perfect. Whereas 80 percent of the individuals over 65 receive the bene t, the share is slightly lower as one approaches the cuto , a situation that is mainly driven by the indigenous male population in rural areas. Low take-up rates are somewhat unsurprising given that, in rural areas, there are signi cant de ciencies in personal identi cation systems and there are few nancial centers, a situation that tends to dispro- portionately a ect the indigenous population 4 . Under the reasonable assumption that, on average, age is truthfully reported and that it is measured with little error, it is reassuring to observe that very few elders who are ineligible to bene t from this program are cashing out of Bolivida. 2 Furthermore, the ethnic dimension is closely correlated to poverty as 49 percent of indigenous people live below the poverty line, but only 24 percent of the non-indigenous population do (http://www.ine.gob.bo). 3 As in many Latin American countries, in the 1990s Bolivia implemented a broad array of structural reforms, which included the sell-o of state-owned companies to private investors and the change of the pension system from a public pay-as-you-go system to a privately managed one. 4 There are no data that allow assessing the extent to which di erent constraining mechanisms might be deterring elders to cash out of Bolivida. 5 3 Data We employ a set of nationally representative cross-sectional Living Standards Measurement Surveys (LSMS) from 1999 and 2002 administered by the Bolivian National Institute of Statistics5 . The surveys include a comprehensive socioeconomic module, Bolivida receipt information at the individual level, and detailed expenditures for all members in the household who are at least six years old6 . The sample comprises all school-age children who live in households with at least one person in the age range of 56 to 737 . The school-age range considered is 6- to 18-years as the minimum legal working age in Bolivia is 14. The sample excludes households that belong to the top 1 percent of the educational expenditures distributions. The total sample includes 3,645 school-aged children and 1,038 eligible elders distributed among 1,915 households. The surveys employ a strati ed two-stage sampling whose sampling frames are based on census data8 . Table 1 reports summary statistics. On average, families living with an eligible elder are not substantially di erent from those living with a soon-to-be eligible elder, factors that have not changed signi cantly during our period of study. Eligible households are slightly smaller and have slightly more children than non-eligible ones, even though both have nearly the same number of school-age children, especially in later years. However, non-eligible and eligible households have di erent family characteristics. Particularly, a higher percentage of non-eligible households have both the father and mother living in the household, and they are of parental age. In addition, the parents’ years of schooling and the children’s education expenditures are higher for eligible households. Finally, both eligible and non-eligible families allocate approximately 10 percent of their total income to children’s educational expenditures9 . 5 INE website: http://www.ine.gob.bo. 6 Approximately 18 percent of households are three-generation extended families. 7 The preferred sample is de ned based on the bandwidth used for the estimation of the regression discontinuity model. Robustness tests were conducted using slightly larger and smaller samples. 8 The sampling frame for the baseline survey was constructed based on the 1992 census enumeration areas list. The follow-up survey uses an updated sampling frame that was constructed upon revised cartographic information compiled for the 2001 census. While this may raise questions regarding the bias of the estimates, we perform robustness tests that show no statistical e ect of this sampling framework. These results are available upon request. 9 We also observe that the largest disparities appear in human capital among ethnic groups as well as between urban and rural populations where educational gender gaps among indigenous people are particularly large. 6 To identify the e ects of the unconditional cash transfer on expenditures in education, we build a variable that captures expenditures on education for each school age child as our key unit of analysis. Expenditures on education include education enrollment tuition and fees, uniforms, school materials, other school fees and contributions, transportation related expenditures and individual expenses for each student10 . This measure captures heterogeneity across di erent age cohorts, as expenditures may di er across educational levels, and intra-household human capital investment decisions may vary at di erent stages of life. Therefore, to evaluate the impact on education investment, our unit of analysis is children between 6 and 18 years old11 . Accordingly, we analyze the combined impact on the decision to invest, if the children have matriculated, and the amount spent. In terms of treatment assignment, we identify the treatment status of a household based on the age of the household member closest to the cut-o age of 65 years. If the household member closest to the cut-o age is younger than 65, the household is considered untreated, and if the member is 65 years old or older, the household is considered treated. For example, a household with two senior people, one who is 64 and one who is 80, is considered untreated, while a household with two senior persons, one who is 65 and one who is 80, is considered treated. In this sense, the results can be interpreted as the impact of having one additional recipient member, and we can more clearly identify the impact of receiving the cash transfer on the allocation of expenditures within the household. Thus, we consider as treated those children who live in a treated household and consider as untreated those children living in an untreated household. Panel (a) in Figure 2 displays a histogram of individuals at age 50 and above that compares before and after the implementation of Bolivida, which occurred at the end of 2000. There is no evidence of a sudden increase in the population distribution at the age of the threshold of eligibility. Panel (b) in the same gure indicates that there is no change in the share of households that are eligible around the cut-o age. In addition, the conditional probability of being eligible increases at such a 10 For additional details see the household questionnaire for the 2001 Household survey, section 9 part E: Education expenditures. We exclude from the calculations expenditures on photocopies due to inconsistent information reported and outliers in the sample. 11 Additional estimations using household per-student educational expenditure as well as including children in school but without expenditures were performed and the results remain robust. 7 value, indicating a signi cant discontinuity with respect to actual bene ciaries. As it is reasonable to believe that age cannot be manipulated, especially since bene ciaries are required to provide legal documents to authorities to participate, households are unlikely to strategically locate themselves around the age of eligibility. Thus, we can con dently assert that the main assumption behind our regression discontinuity design holds, and therefore, we use it to identify the average UCT intention- to-treat e ects on children’s schooling expenditures (Imbens and Lemieux, 2008)12 . It is noted that we do not estimate the average treatment e ect on the treated as the use of the bene ciary variable in this setup is problematic since the receipt of a pension may be endogenous given that we nd slight di erences between actual and estimated bene ciaries. As previously mentioned, these di erences may arise because of de ciencies in the government’s personal identi cation documentation system, because of a lack of available funds to cover the fully targeted population, or because of di culties reaching nancial centers, either due to the lack of transportation or the excessive distance. All these factors may constrain eligible members from receiving bene ts13 . Figure 3 presents the conditional expectation of children’s schooling expenditures as a function of the age of the household member closest to the cut-o age. The smoothing is performed sepa- rately for the period before and the period after 2001. The graph makes clear that before 2001, there is no di erence in the average education expenditure per child. After 2001, however, there is an im- portant and statistically signi cant increase, which signals that regression discontinuity methods are appropriate. Next, we parameterize these relationships to quantify the e ect of Bolivida on children’s educational expenditures. 4 Empirical Strategy The aim of this paper is to analyze the causal e ect of the potential unconditional cash transfer re- ceived by elders on the education expenditures of children who live in the same household, when 12 The density of age around the cut-o appears symmetric, which supports the validity of the assumption. 13 As the exact constraining mechanisms are unclear, we decided that our empirical exercises should focus mainly on potential bene ciaries. 8 taking advantage of the exogenous implementation of the Bolivida program of 2001, which is related to the income shock that families with an elder member may have experienced. Based on the char- acteristics of the program, a possible strategy may be to adopt a di erences-in-di erences approach, comparing education expenditures per child as a function of the age of eligibility of the reference household member and the variable indicating whether Bolivida was already implemented. How- ever, a simple comparison of education expenditures by households below or above the cut-o age across time is likely to provide a biased estimate. Since the cut-o age coincides with the retirement age, the household structure and distribution of potential retirees may not be random, and thus, the presence of elder household members may have an impact on the allocation of time and resources among household members. Therefore, it may be possible that potential bene ciary households will di er systematically from non-bene ciary households. In this case, a simple di erence-in-di erence model would not be valid to account for the potential systematic di erences. To overcome unobserved di erences between potential bene ciary and non-bene ciary house- holds, we exploit the fact that the probability that an individual receiving the cash transfer changes discontinuously at the age of eligibility. Accordingly, we apply a straightforward regression discon- tinuity design (Imbens and Lemieux, 2008). Furthermore, since we also have information across time, namely, individual information before and after the implementation of the program, an additional ap- proach that we apply is a combination of a di erences-in-di erences approach and a sharp regression discontinuity design. This speci cation is written as follows: log (Educ exp + 1) =a0 + a1 ∗ Iage65 + a2 ∗ Iyr≥01 + a3 ∗ Iage65 ∗ Iyr≥01 + bγ ∗ fγ (agereg − 64.5) + bl ∗ Iage65 ∗ fl (ager eg − 64.5)+ (4.1) cγ ∗ Iyr≥01 ∗ gγ (ageref − 64.5)+ cl ∗ Iyr≥01 ∗ Iage65 ∗ gl (ageref − 64.5) + where the log of per child education expenditure is estimated as a function of the age of eligibility 9 (Iage65 ) of the reference person in the household, an indicator of the year when the transfer was implemented (Iyr≥01 ) and an interaction of both age of eligibility and year of implementation. In addition, we allow for the inclusion of exible functional forms of the gap between the reference person’s age and the cut-o age (ageref -64.5), both of which can vary before and after the year of implementation (fγ ,fl ,gγ ,gl ), and for households with a reference individual’s age above or below the cut-o age. In this speci cation, a3 is the parameter of interest as it identi es the impact of the unconditional cash transfer for school age children living in households where the reference person is near the cut-o age14 . Since the treatment state of a household is based on the age of the person closest to the cut- o age or the reference individual, it is possible that there may be other elderly individuals living in the household. Given that their presence is also likely to in uence household expenditures on children possibly due to income shocks, our baseline speci cation includes a control for the presence of any household member older than 65, other than the reference person. In the context of the regression discontinuity approach, two additional factors are considered, (i) the functional form for the age gap (fγ ,fl ,gγ ,gl ) and (ii) the bandwidth selection. With respect to the choice of the functional form, we consider that the forcing variable, the age of the reference individual, is discreet, and thus, the use of non-parametric regressions is not viable. Instead, we estimate the empirical speci cations using di erent polynomial functions, including logistic, linear, quadratic, and cubic functional forms. Due to the restricted variation in the forcing variable, we do not estimate higher-order polynomials, which helps us avoid over-speci cation. The selection of the order polynomial is based on the Akaike information criteria (AIC) as suggested by Lee and Lemieux (2010). The second key issue is the choice of an appropriate bandwidth selection, as it determines the nal sample employed in the analysis. While larger bandwidths allow for the use of more data and, thus, potentially help obtain more precise estimates, they can also reduce the accuracy of the estimates if the model is misspeci ed, which may introduce bias on the estimated treatment e ects. To address 14 As indicated in Lee and Lemieux (2010), the estimation of the causal e ect can be tested using separate regressions for samples above and below the cut-o . While we provide results based on the pooled regression speci cation, we also provide results based on the estimations of separate regressions in Appendix 1 (Table 9). 10 this issue, we apply a cross-validation strategy, as described by Lee and Lemieux (2010), to select the appropriate bandwidth. Based on the cross-validation strategy and the AIC applied to the baseline model, we choose a linear model with a bandwidth of nine years above and below the cut-o age. 5 Results We depart from testing competing models to obtain an initial estimate as well as establish the best speci cation for our estimation. As evidenced in Figure 5, and following the existing literature of Lee and Lemieux (2010), we test quadratic, cubic, log linear and linear speci cations. For most speci ca- tions, the average e ect is below one. Following Lee and Lemieux (2010) and the Akaike criteria, these ndings indicate that the preferred empirical speci cation should be the linear speci cation. Table 2 reports the basic results of our preferred speci cation without any controls. Column one shows that, on average, the implementation of the cash transfer increased expenditures on education by approxi- mately 73 percent in Bolivia. In addition, our results show that the probability of moving from a zero expenditure on education to a positive size is approximately 12 percent and, for those households that have some level of expenditure, the e ect of the cash transfer reaches approximately 43 percent and remains statistically signi cant at conventional levels. These results indicate that unconditional cash transfer to elders in uences intra-household behavior by promoting larger expenditures on ed- ucation. Moreover, these e ects are larger in those households that are already contributing to their children’s education. However, these results may be biased as we are not controlling for other factors that may a ect education expenditures. Table 3 presents the ndings when we include a broad set of controls, speci cally, age, gender, number of children in the household, number of children aged ve or under, number of adults, pres- ence of a parent in the household, parents’ education, urban or rural area, and indigenous or non- indigenous background. For the most part, we nd analogous results and stronger impacts. However, it is interesting to note that when controlling for an indigenous background and an urban setting, the impact observed is reduced by approximately ten percentage points with respect to the speci cation 11 without such controls. These somehow suggest that additional heterogeneity analyses can provide further insights into the results presented herein. For instance, full sample estimates may hide some heterogeneity due to cultural and geographical factors, and, more particularly, ethnicity, a highly relevant element in a country such as Bolivia. The identi cation of indigenous people in Bolivia is not straightforward as social class and ethnic elements are interrelated and di cult to disentangle, and, in general, the information used to de ne who the indigenous people are is based on a set of ethnolinguistic characteristics. We use three speci c questions from our survey data aimed at identifying ethnic groups: (i) do you consider yourself as belonging to an indigenous group? (ii) what languages do you speak? and, (iii) what language did you rst learn to speak? The rst two questions are intended exclusively for household members who are at least 12-years-old, and the third is intended for all members in the household. We follow Molina and Albó (2006), who employ the above three criteria to construct an ethnolinguistic matrix for Bolivia. We classify as non-indigenous those who do not self-identify as indigenous, whose rst language was not indigenous, and who did not learn an indigenous language as children. We de ne as indigenous those who identify themselves as indigenous, whose rst lan- guage was indigenous and those who learned an indigenous language as children. To simplify our empirical approach, all other combinations are labeled as multiethnic. Table 4 displays the heterogeneous impacts of the unconditional cash transfer of elders on chil- dren’s educational expenditures by ethnicity, region, gender, and age cohorts. The estimated param- eters are reported separately for each group. After the implementation of the cash transfer, we nd statistically signi cant and larger impacts for indigenous populations and smaller and statistically in- signi cant e ects for non-indigenous people. Mixed race populations reveal statistically insigni cant results that lie somewhere between the indigenous and non-indigenous populations. These results may be because indigenous populations tend to face credit constraints, and therefore, any liquidity shock modi es signi cantly the allocation of expenditures within households. A similar nding is noted when comparing rural and urban groups. While in the former we nd a large and statistically signi cant impact, urban populations exhibit smaller and non-signi cant e ects. In addition, we es- 12 timate the average e ects of eligibility for girls and boys separately15 . The coe cients are reported in column (4) of Table 4. We nd that conditional on eligibility, human capital expenditures on girls almost doubles and is strongly statistically signi cant, while for boys, this result is approximately 22 percent and not statistically signi cant at any standard levels. This nding is reassuring since females have been historically neglected in schools in Bolivia. In Table 5, we further explore the role played by gender in the intra-household allocation from elders to children. Speci cally, we explore the a nity between the gender of grandparents and grand- children as well as the impact of the age of the children on grandparents’ preferences. We nd ev- idence that educational expenditures on girls are driven by grandfathers, but not by grandmothers, as revealed in Panel (a). This is rather remarkable, as it indicates that male elders, having bene tted from a system that favors males, realize the di culties that females have in Bolivia. Moreover, it is fully consistent with our previous ndings. Furthermore, we nd that grandfathers focus their attention on girls between the ages 13 and 18 years, unlike grandmothers, who appear to focus on younger children. This is presented in Panel (b) in Table 5. 6 Robustness We observe that there is a small share of the population that may be receiving the unconditional cash transfer despite being younger than the corresponding cut-o age of 65 and that the share of people 16 who are not receiving the bene t may be greater for those closer to the cut-o . This may raise questions about the strength of using 65 years of age as the cut-o point and about the relevance of the bandwidth. To test for this, we perform a series of falsi cation tests, the results of which are displayed in Table 6. We test alternative cut-o ages ranging from 63 to 67 years of age and control for analogous covariates as in our main results. We do not nd any statistically signi cant results at conventional levels. 15 Child ethnicity sub-samples cannot be tested since not all variables used in the classi cation of ethnic groups were collected for children under the age of 12. 16 We observe this graphically. The corresponding gure is available upon request. 13 Similarly, Table 7 presents the results using di erent bandwidth and alternative empirical speci- cations that control for all the covariates as in Table 6. A major concern with the regression discon- tinuity design is the high sensitivity of the estimates to the choice of bandwidth, and thus, although moving the bandwidth slightly a ects the results, they remain stable and statistically signi cant for our preferred model, the linear speci cation. The quadratic and cubic models provide similar es- timated treatment e ects, albeit smaller for the quadratic and larger for cubic. However, for the quadratic speci cation, it is only signi cant when using h = 10, and it is not statistically signi cant in the cubic case As an additional robustness test, Table 8 reports ndings that are not estimated using a regression discontinuity design along with a di erences-in-di erences approach but rather are estimated using a combination of RDD and a before-after design (Hoddinott and Skou as, 2004; Borooah and Iyer, 2005). That is, these results are estimated separately for the period before and the period after the Bolivida UCT, and conditional means are then compared to assess the impact of the program. This speci cation is more exible than the pooled one previously employed, but it reduces the degrees of freedom as it requires that more parameters be estimated. Although the results are somewhat smaller, they are equally statistically signi cant. Finally, the Appendix (Table 9) presents the results when using weights provided by the National Statistics Institute. If anything, the results tend to be more robust and slightly greater than, albeit not signi cantly di erent from, our base results17 . 7 Conclusions In this research, we take advantage of repeated cross-section household surveys and a sharp dis- continuity created by the introduction of an unconditional cash transfer to elders to evaluate the 17 The household survey data collection process employs a two-stage sampling strategy. First, primary sampling units are chosen, and then, households are selected randomly. This may call for the use of weighting schemes, which are estimated by the National Statistical O ce. Since our speci cation is based on a pooled cross-section, a di -in-di approach may yield biased estimates. This is because the sampling frame for the baseline surveys (1999 and 2000) was constructed based on the 1992 census enumeration list while the follow up surveys (2001 and 2002) used an updated sampling frame that was constructed upon revised cartographic information compiled for the 2001 census. 14 intent-to-treat impact on the educational expenditures related to children within a household by us- ing a regression discontinuity design and a di erences-in-di erences approach. The eligibility e ect, which we allow to vary across households and children, is assumed to be a linear function of the ethnicity and gender of the potential recipient. The main nding is that unconditional cash transfers to elders lead to substantial improvements in children’s human capital investments. We also nd that the program has stronger e ects in indigenous populations as well as in female and rural popu- lations. Our results are robust to a series of falsi cation tests, survey structures, model speci cations and estimation methods. References Agüero JM, Carter MR, Woolard I. 2005. From ows to stocks: The impact of unconditional cash transfers on human capital. Working Paper, University of Conneticuc . Amuedo-Dorantes C, Juarez L. 2015. Old-age government transfers and the crowding out of private gifts: The 70 and above program for the rural elderly in mexico. Southern Economic Journal 81: 782–802. Azariadis C. 1993. Intertemporal macroeconomics . Baird S, McIntosh C, Özler B. 2011. Cash or condition? evidence from a cash transfer experiment. The Quarterly Journal of Economics : qjr032. Barrientos A. 2003. What is the impact of non-contributory pensions on poverty? estimates from brazil and south africa. Working paper, University of Manchester . Barro RJ. 2004. Sala-i martin x (2004) economic growth. Au . Massachusetts Institute of Technology (MIT) ISBN: 0-262-02553-1 . Behrman JR. 2011. How much might human capital policies a ect earnings inequalities and poverty? 15 Estudios de Economia 38: 8–41. URL https://ideas.repec.org/a/udc/esteco/v38y2011i1p8-41.html Borooah VK, Iyer S. 2005. Vidya, veda, and varna: The in uence of religion and caste on education in rural india. The Journal of Development Studies 41: 1369–1404. Burlando A. 2014. Transitory shocks and birth weights: Evidence from a blackout in zanzibar. Journal of Development Economics 108: 154–168. Du o E. 2003. Grandmothers and granddaughters: old-age pensions and intrahousehold allocation in south africa. The World Bank Economic Review 17: 1–25. Edmonds EV. 2006. Child labor and schooling responses to anticipated income in south africa. Journal of development Economics 81: 386–414. Escobar-Loza F, Martinez-Wilde S, Mendizábal-Córdova J. 2013. El impacto de la renta dignidad: polıtica de redistribución del ingreso, consumo y reducción de la pobreza en hogares con personas adultas mayores. Bolivia: Unidad de Análisis de Polıticas Sociales y Económicas del Ministerio de Plani cación del Desarrollo, Representación en Bolivia del Fondo de Población de las Naciones Unidas & HelpAge International . Hernani-Limarino W, Mena G. 2015. Intended and unintended e ects of unconditional cash transfers: The case of bolivia’s renta dignidad. Technical report, IDB Working Paper Series. Hoddinott J, Skou as E. 2004. The impact of progresa on food consumption. Economic development and cultural change 53: 37–61. Imbens GW, Lemieux T. 2008. Regression discontinuity designs: A guide to practice. Journal of Econometrics 142: 615–635. URL https://ideas.repec.org/a/eee/econom/v142y2008i2p615-635. html 16 Joubert C, Todd P. 2011. The impact of chile’s 2008 pension reform on labor force participation, pension savings, and gender equity. Univ. of Pennsylvania, Working Paper . Landa F, Yanez-Pagans P. 2008. Informe especial: La informalidad en el mercado laboral urbano: 1996-2006. Manuscript, World Bank . Lee DS, Lemieux T. 2010. Regression discontinuity designs in economics. Journal of economic litera- ture 48: 281–355. Martinez S. 2004. Pensions, poverty and household investments in bolivia. Department of Economics, University of California at Berkeley . Molina R, Albó X. 2006. Gama étnica y linguística de la población boliviana. Technical report, Sistema de las Naciones Unidas en Bolivia. The World Bank. 2016. World Development Indicators. www.worldbank.org. Valencia H. 2011. Efecto del ingreso en las decisiones de co-residencia de los adultos mayores en bolivia. Tesis, Magíster en Economía, Instituto de Economía - Ponti cia Universidad Católica de Chile . 17 Table 1: Summary Statistics Non-Eligible (Age ≤ 64) Eligible (Age≥65) 1999-2000 2001-2002 1999-2000 2001-2002 Total Panel A: Demographics Child Age 12.5 12.6 12.1 11.8 12.4 Share Men (percentage) 50.3 51.8 54.9 48.2 51.0 # School Age Children in HH 2.6 2.6 2.3 2.8 2.6 # Children under 5 in HH 0.5 0.6 0.6 0.6 0.6 Household Size 6.0 6.1 5.8 6.3 6.1 # Adults 19-64 2.8 2.9 1.9 1.9 2.6 Father Present 0.7 0.7 0.6 0.5 0.7 Mother Present 0.8 0.8 0.7 0.7 0.8 Panel B: Parents Yrs of Education Father 5.2 4.8 6.0 5.3 5.1 Yrs of Education Mother 3.9 4.0 5.4 5.1 4.3 Age Father 54.5 54.3 51.9 51.4 53.6 Age Mother 49.2 49.3 47.6 47.1 48.7 Panel C: Income, Expenditure (in Bs. Per month) Income per capita w/o Bono 478.1 505.7 432.5 501.6 489.1 Expenditure per capita 445.1 396.6 495.8 451.9 431.2 % Children with Positive expenditure in education 78.5 77.1 82.9 83.4 79.3 Expenditure on education per child 51.6 34.5 62.1 46.2 44.5 Observations 1,067 1,540 386 652 3,645 Table 2: Impact of UCTs on Educational Expenditures (1) (2) (3) Log (Expenditure+1) Pr(Expenditure>0) Log Expenditure if Expenditure > 0 Treatment 0.732** 0.123* 0.426* (0.255) (0.059) (0.203) N 3645 3645 2889 R2 0.025 0.014 0.031 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 18 Table 3: Impact of UCTs on Education Expenditures with Controls 1 2 3 4 5 6 7 8 9 10 11 Treatment 0.732** 0.667*** 0.607*** 0.573*** 0.601*** 0.629*** 0.612*** 0.633*** 0.637*** 0.636*** 0.586*** -0.255 -0.224 -0.221 -0.22 -0.219 -0.219 -0.219 -0.219 -0.219 -0.218 -0.213 Urban × Dept HH Indig Age Gender 19 Children HH # Children<5 # Adults Parent Present Parent > 65 Parent educ N 3645 3645 3645 3645 3645 3645 3645 3645 3645 3645 3645 R2 0.025 0.216 0.232 0.261 0.263 0.271 0.274 0.275 0.279 0.281 0.329 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 Table 4: Heterogeneous Impacts (1) (2) (3) (4) Indigenous Status Urban Status Age Gender Non Indigenous 0.232 (0.350) Observations 1480 R2 0.355 Indigenous 0.727** (0.370) Observations 1157 R2 0.261 Mix 0.553 (0.450) Observations 1008 R2 0.300 Rural 0.746** (0.309) Observations 1809 R2 0.204 Urban 0.434 (0.289) Observations 1836 R2 0.238 6-12 0.612** (0.266) Observations 1799 R2 0.331 13-18 0.462 (0.337) Observations 1846 R2 0.352 Girls 0.966*** (0.309) Observations 1785 R2 0.354 Boys 0.225 (0.295) Observations 1860 R2 0.321 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 20 Table 5: Characteristics of Elder and Child A. Gender of Child and Gender of Elder (1) (2) (3) (4) Girl/Grandmother Girl/Grandfather Boy/Grandmother Boy/Grandfather Treatment 0.566 1.484*** 0.300 -0.005 (0.469) (0.429) (0.489) (0.388) Observations 756 1029 749 1111 R2 0.388 0.368 0.330 0.346 21 B. Gender of Elder and Age of Child (1) (2) (3) (4) Grandmother/6-12 Grandfather/6-12 Grandmother/13-18 Grandfather/13-18 Treatment 0.800** 0.486 -0.116 0.734* (0.389) (0.385) (0.580) (0.422) Observations 758 1041 747 1099 R2 0.364 0.345 0.369 0.373 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 Table 6: Falsi cation Tests (1) (2) (3) (4) Age cut=63 Age cut=64 Age cut=66 Age cut=67 Treatment -0.042 0.112 0.295 0.213 (0.191) (0.199) (0.232) (0.237) Observations 4436 4098 3225 2988 R2 0.336 0.338 0.330 0.335 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 Table 7: Sensitivity to Bandwidth and Speci cations h=7 h=8 h=9 h=10 h=11 Linear 0.447* 0.507** 0.586*** 0.489** 0.451** (0.241) (0.229) (0.213) (0.200) (0.190) N 2696 3100 3645 4223 4728 R2 0.345 0.334 0.329 0.336 0.330 Quadratic 0.513 0.434 0.426 0.679** 0.702** (0.347) (0.326) (0.306) (0.290) (0.278) N 2696 3100 3645 4223 4728 R2 0.346 0.334 0.329 0.336 0.331 Cubic 0.585 0.656 0.654 0.397 0.516 (0.513) (0.462) (0.421) (0.392) (0.369) N 2696 3100 3645 4223 4728 R2 0.347 0.335 0.329 0.336 0.331 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 22 Table 8: Regression Discontinuity with Before and After Approach 1 2 3 4 5 6 7 8 9 10 Treatment 0.618*** 0.550** 0.520** 0.542** 0.571*** 0.541** 0.578*** 0.586*** 0.528** 0.446* (0.222) (0.219) (0.218) (0.217) (0.217) (0.217) (0.222) (0.221) (0.238) (0.228) Urban × Dept HH Indigenous status Age Gender 23 # Children in HH # Children under 5 # Adults Parent present Father/Mother > 65 Father/Mother educ Observations 3645 3645 3645 3645 3645 3645 3645 3645 3645 3645 R2 0.242 0.254 0.283 0.286 0.291 0.293 0.296 0.301 0.302 0.356 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001 Local polynomial smooth 1 Share of Receiving Individuals .2 .4 0 .6 .8 40 45 50 55 60 65 70 75 80 85 Age kernel = epanechnikov, degree = 0, bandwidth = .72, pwidth = 1.08 Figure 1: Share of People who Received UCT by Age 1999/2000 2001/2002 .1 Density .05 0 50 60 70 80 50 60 70 80 Age Figure 2: Histogram in Age Person Closest to Eligibility 24 .7 .6 .5 .4 .3 55 60 65 70 75 Age 95% CI 65 - 75 years 55 - 65 years Figure 3: Share of Households eligible around the Threshold Age 3.5 3 2.8 3 2.6 2.4 2.5 2.2 2 2 55 60 65 70 75 55 60 65 70 75 Age Age (a) (b) Figure 4: Average Log Expenditure on Education per children (1999/2000) 25 Linear Model Quadratic Model 4 4 3 3 2 2 1 1 0 0 0 5 10 15 0 5 10 15 Bandwidth Bandwidth CI Estimated Treatment CI Estimated Treatment Cubic Model Log Linear Model 4 4 3 3 2 2 1 1 0 0 0 5 10 15 0 5 10 15 Bandwidth Bandwidth CI Estimated Treatment CI Estimated Treatment Figure 5: Base Model Polynomial Estimations 26 Table 9: Appendix: Results Using Weights (1) (2) (3) (4) (5) (6) (7) (8) (9) (10) (11) Treatment 0.766*** 0.672** 0.663** 0.688** 0.738*** 0.738*** 0.771*** 0.751*** 0.749*** 0.749*** 0.665** (0.277) (0.278) (0.274) (0.275) (0.275) (0.274) (0.275) (0.271) (0.270) (0.263) (0.261) Urban × Dept HH indigenous Age Gender 27 # Children in HH # Children under 5 # Adults Parent present Father/Mother > 65 Parent educ Log Househol Inc N 3645 3645 3645 3645 3645 3645 3645 3645 3645 3645 3645 R2 0.242 0.254 0.283 0.286 0.291 0.293 0.296 0.301 0.302 0.356 0.363 Standard errors in parentheses, * p < 0.05, ** p < 0.01, *** p < 0.001