Non-Contributory Pensions and Households Decisions. Evidence from Bolivia

Transcription

1 Non-Contributory Pensions and Households Decisions. Evidence from Bolivia Miguel Angel Borrella Mariano Bosch Marcello Sartarelli October, 25 Abstract We study empirically the effect of a universal non-contributory pension in Bolivia, Renta Dignidad, on households income, labour supply, consumption and living arrangements decisions. We estimate it, accounting for the gender of the spouse receiving a pension, thanks to rich survey data and a regression discontinuity design, as individuals pension eligibility jumps discontinuously at age 6. We find that, despite increasing disposable income, receiving one or more pensions has a small and not significant impact on poverty levels in old age at the household level. This is due to the fact that receiving two pensions triggers an increase in household size as extra grandchildren are observed in beneficiary households, while a pension only to the husband (de)increases transfers (from)to other households. Failing to account for decisions taken at household level and for beneficiary s gender would lead to mistakenly conclude that the pension has no effect on labour supply and living arrangements. Thanks to our novel research design, instead, we uncover rich behavioural response to the pension, suggesting that, although it targets individuals, important decisions are taken at the household level, may be gender-specific and also indirectly affect other household members as well as other households. JEL Classification: H2, H3, J2 We thank for helpful comments Fernanda Brollo, Lola Collado, Giacomo De Giorgi, Giacomo Deluca, Tommaso Nannicini, and participants in: seminars at U. Alicante and at the Tinbergen Institute in Amsterdam, the 23 EUI Alumni Conference, the 23 Econometric Society Meeting in Gothenburg, the IZA Workshop on Labor Markets and Labor Market Policies for Older Workers and the 24 Petralia Sottana Applied Economics Workshop. Financial support from the Spanish Ministerio de Economía y Competitividad (ECO and ECO23-439) and from the Universidad de Alicante (GRE 3-4) is gratefully acknowledged. All remaining errors are our own. Universidad de Alicante. Corresponding author: Inter American Development Bank, 3 New York Avenue N.W., Washington, D.C. 2577, USA, mbosch@iadb.org

2 Keywords: Consumption, labour supply, living arrangements, noncontributory pension, regression discontinuity, Renta Dignidad, transfer

3 Introduction Traditional Bismarkian pension systems, based on contributions during workers lives to finance old age pensions, have failed to provide coverage to significant parts of the population in low and middle income countries. This is the result of lack of compliance with tax and pension systems due to widespread informal unregulated jobs, sometimes called informal jobs. A century after the introduction of pension systems around the world, only 45%, 35% and less than % of the workers contribute to a pension system in Latin America, Asia Pacific and Africa, respectively (World Bank, 25). As a response to the lack of coverage and in the face of rapidly changing demographics, countries are moving to provide pensions through noncontributory or social pensions. These systems provide income for elderly population regardless of their employment or social security contribution history. Over the last 2 years, 6 countries have adopted this arrangement to prevent old age poverty (HelpAge, 25). Even in relatively young countries, like Bolivia, Brazil or Chile, expenditure for these programs absorbs more than % of GDP and, given the demographic transition, it is likely to increase (Bosch et al., 23). The advent of non-contributory pensions is bound to change household decisions, particularly in poor countries, where the expectation for the elderly was either to keep working indefinitely or to rely on family members for their subsistence. Understanding those changes is crucial for policy design to avert old age poverty in a context of rapid aging in low and middle income countries. In this paper we study the effect of a universal non-contributory pension in Bolivia (Renta Dignidad). Bolivia is one of the poorest countries in Latin America and a country where the contributory pension system only provides coverage to around % of the elderly population (Bosch et al., 23). Renta Dignidad was introduced in 28 with the implicit aim of reducing elderly poverty. It consisted on a monthly transfer to all citizens aged 6 or older of 2 bolivianos. This corresponds to 25% of per capita See Acosta et al. (24); Bosch and Manacorda (22); Bosch et al. (23); Mesa- Lago (24); Rofman et al. (28, 24) for recent assessments of pensions policies in Latin America and the Caribbean region. See instead Holzmann and Hinz (25) for a broader assessment of pensions policies in 6 developing countries. 2

4 income and about 6 US dollars in PPP (Bosch et al., 23). Using a regression discontinuity design (RDD), as pension eligibility jumps discontinuously when an individual turns 6, we study whether and how household decisions were reshaped soon after one or both spouses became eligible for it. We estimate a fuzzy RDD thanks to Two Stages Least Squares, as take-up is 6-7%, by using pension eligibility as an instrument for take-up. In addition, thanks to variation in spouses age in a household, we can estimate the effect by gender of the beneficiary using spouses age as running variables and eligibility by spouse as instruments. We find significant changes in the way households generate income and in their living arrangements. Combining these significant short-term behavioural responses with a relatively low pension leads to a negative although not significant impact on extreme poverty rates, by about percentage points, pp hereafter, or 3% with respect to households with no pension, in parenthesis hereafter. Interestingly, Renta Dignidad targeted individuals and not households although of all households in which at least a spouse is eligible for a pension, both spouses are in 38% of them. Most of our results are driven from this group that weakly reduces female labour supply by 3 pp (5%) and for whom the probability of absorbing additional family members increases: grandchildren by 5 pp (268%) and sons living and working in the household by 55 pp (3%). When, instead, only the husband receives it, transfers from other households decrease by 236 bolivianos (245%) and those sent to other households increase by 98 bolivianos (35%). When any spouse obtains the pension, instead, the probability of purchasing medicines increases by 65 pp (4%). Overall, most results also hold when testing for the pension effect jointly on several outcomes. Finally, by failing to account for the gender of the beneficiary, we would mistakenly conclude that the only pension effect is on medicine purchase and, similarly, that the pension does not affect decisions had we also not accounted for the fact that certain decisions may be taken at the household level. Importantly for our research design validity, we find that the distribution of spouses age and baseline characteristics are continuous at the 6 cutoff, which rules out sorting by individuals at 6. In addition, falsification tests obtained estimating the age 6 effect by using data in the period 3

5 26-27, i.e. before the pension was enacted, exclude that the results are driven by a spurious demographic effect at 6. Finally, results are robust to varying the time elapsed between when the pension is first received and when outcomes are observed, the RDD bandwidth and also to dropping individuals eligible for the Bonosol pension, that preceded Renta Dignidad and whose cutoff age was 65. Although we are not the first to study the effect of a non-contributory pension on the well-being of its beneficiaries, some authors focused primarily on psychological tests as proxies for it. Galiani et al. (24) show that mental health of elderly adults is significantly improved after the introduction of non-contributory pension in rural Mexico while, in contrast, Novella and Olivera (24) find a negative effect of a similar pension in Peru on cognitive ability. This highlights the importance of obtaining additional evidence on the mechanism underlying these results. Focusing, instead, on indirect effects, Duflo (23) shows that the expansion of a pension in South Africa had positive effects on child health. By exploiting similar non-contributory pensions enacted before Renta Dignidad in Bolivia, Yanez-Pagans (28) and Martinez (24) find a positive effect on children s schooling. We extend this literature by documenting thanks to what set of related household decisions taken jointly by beneficiaries their well-being and that of other household members may increase. In addition, we assess whether pension gender effects and receiving a pension per spouse, also influence their choices. Perhaps the most widely documented result, in several countries, is that having access to a non-contributory pension seriously impacts the labour supply of beneficiaries. In line with theoretical predictions, the pension significantly reduces it (de Carvalho Filho, 28; Bosch and Guajardo, 22; Kassouf and Rodrigues de Oliveira, 22; González and Pfutze, 24). Evidence on how the pension affects household living arrangements is, instead, less clear-cut as additional family members may cluster around the constant stream of pension income (Duflo, 23; Hernani-Limarino and Mena, 25; Yanez-Pagans, 28) or, in contrast, may leave the household (Hamoudi and Thomas, 24; Posel et al., 26). We extend this literature by studying whether and how the gender of a beneficiary affects household decisions rather than individuals decisions. 4

6 Finally, if private transfers are crowded out by a pension, the beneficiary is worse off as the intended effect at the policy design stage is, in fact, diluted and vice versa if it is crowded in. Juarez (29) and Amuedo-Dorantes and Juarez (23) find that a non-contributory pension to individuals turning 7 or older in Mexico decreases them by about 3% in urban areas and %, i.e. complete crowding out, in rural areas. Similarly, Jensen (24) finds a decrease by about 3% in South Africa. In related work on Progresa transfers in Mexico, Albarran and Attanasio (23) find that the magnitude of crowding out decreases with village-level variance in income, in line with models relating risk-sharing and limited contract enforceability. Exploiting the same program, Angelucci et al. (22) find evidence consistent with transfers to other households as it increases consumption for ineligible households belonging to the family network of eligibile ones. We contribute to this growing literature with novel evidence that crowding out is gender-specific and, in addition, it may be observed along with pensioninduced transfers to other households. This suggests a novel mechanism to explain variation in inter-households transfers, an example of informal contract with limited enforceability. Overall, this paper makes three contributions to the literature and to inform policy decisions. First, we exploit a novel source of variation, combining the universality of the program with a common age eligibility cutoff and variation in couples age, to study behavioural responses not only to variation in pension eligibility across households but also in the gender of the beneficiary. Second, we show how, even in the very short-run, the pension generates important changes in household decisions that proxy a welfare increase, such as an increase in employment opportunities for young adults in beneficiary households, a decrease in labour supply for elderly and absorption of extra family members. Finally, we show that an important part of the pension is diluted by way of transfers across households in the community. A dollar increase in the pension reduces transfers received by and increase transfers sent by. This is crucial for the design of non-contributory pensions as, while they are targeted to the elderly, their benefits seem to spill over to other households. The structure of the rest of the paper is as follows. Section 2 describes the institutional setting and the data. Section 3 discusses the most relevant 5

7 theoretical predictions to the questions we set out to answer. Section 4 describes the empirical research design and section 5 the results and the validity of the research design. Finally, section 6 discusses results and section 7 concludes. 2 Institutional setting and data Before Renta Dignidad the Bolivian government enacted in 997 Bonosol, a pension paying all citizens turning 65,3 bolivianos per year, or about 27% of per capita income and 85% of income of those living in extreme poverty (von Gersdorff, 997), as part of a broader social and economic reform agenda with two main aims. The first was reducing high income inequality in the country, in the top quartile of the distribution of countries worldwide if measured using the GINI index (CIA World Factbook, 24). The second was dealing with the consequences of the high share of informal workers, about 6% (World Bank, 29), who are not entitled to a contributory pension since they did not pay social security. 2 Bonosol was discontinued in 998 as it was judged financially untenable. In 2 the pension, renamed Bolivida, resumed although its amount decreases to about 42 bolivianos in An additional change in the pension paid to individuals who were 65 or older occurred in 22, when Bonosol was restored and payments increased to,8 bolivianos. Finally, Bonosol payments stopped in December Renta Dignidad was enacted on st February 28 with two main differences with respect to previous pensions. First, the age eligibility cutoff decreased to 6. Second, the amount paid increased to 2,4 bolivianos per year, except pensioners or public sector employees who received,8 bolivianos. 4 2 Only % of workers are entitled to a pension in the three middle quintiles of the income distribution (Rofman et al., 28). 3 See Willmore (26, 27) for additional information about pension reforms in Bolivia. 4 Pension were first paid up to a month after turning 6, either at a bank or authorized military enclosure, by showing their ID card and two photocopies. Alternatively, arrangements were in place to obtain it at home. Individuals obtaining 2,4 bolivianos, about 8% of the beneficiaries, could choose the frequency of the payment, either monthly or a lower one while for the others the pension was added automatically to payments that they usually obtained on a monthly basis. Additional information is 6

8 We estimate the pension effect on household income, labour supply, consumption and living arrangements by using two repeated cross-sections of survey data in the years 28 and 29 from the National Statistics Institute in Bolivia (Encuesta de Hogares). The unit of observation in the data is a household and information about all its members is collected either at the individual or at the household level. 5 We use the following measures of income and consumption. We first proxy income by using poverty dummies, defined in the survey using the absolute poverty threshold. 6 We then use, as direct measures of income, total household income and its counterpart in per capita terms. Consumption is measured in the national currency, bolivianos, in a time interval of up to a month before the survey date, for the following items: total food and total non-food consumed in a household, medicines and medical services. 7 Employment status is measured for each household member and it is defined in the survey as a dummy that is equal to if an individual was employed the week before the survey interview and otherwise. We define the following four dummies measuring labour supply within the household: whether each spouse separately had a job the week before the survey, whether at least one of them had and whether both had. Finally, we use a set of dummies to measure the following living arrangements in a household with an elderly couple: whether sons or daughters, i.e. young adults, live with their parents, whether they work and whether grandchildren live with their grandparents. Our treatments of interest are whether each spouse separately or both spouses in a household obtained the the non-contributory pension, with respect to a counterfactual household with no spouse obtaining it. Since survey interviews were held yearly in November and up to a month elapsed available in Escobar Loza et al. (23). 5 The data can be downloaded from 6 Poverty is defined as a dummy equal to if total per capita household income is lower than the poverty threshold (48 bolivianos/month). Extreme poverty is measured in the same way, except using a lower poverty threshold (26 bolivianos/month). Additional details about the survey design are available in 2.pdf. 7 Total food consumption is measured in adult equivalent terms. In particular, we use the OECD scale and we divide the variable by ( +.7 (N adults ) + N children ). See chapter 2 in Haughton and Khandker (29) for additional details. 7

9 between the time at which an individual who turned 6 submitted the application for the pension and the time the pension was paid, we study the short-term effect of the pension by setting the end of September as the date in which we observe if spouses were 6 or older and hence eligible for the pension, as Figure shows. 8 [Figure about HERE] We compare the outcomes of interests between treated and control households by defining as treated those households in which either one or both spouses obtained the pension, since they were eligible for the pension as they were born earlier than 3th September 948. Similarly, the control group is defined as those households in which both spouses were born later than this date. [Table about HERE] Table shows summary statistics of our outcomes of interest and of households baseline characteristics. It is divided into 2 panels. The lefthand side panel shows means for households with spouses in a 4-8 age band and separately for two subgroups defined by whether both spouses were 6 or older within a household, hence eligible for two pensions. In the panel on the right-hand side the table we report the same means but now comparing households in which at least one spouse in a household was 6 or older with the same control group. In both cases, we also assessed whether differences in outcomes between subgroups are significant and report the p-value of the null hypothesis that they are zero. 9 The top panel in the table shows that the pension take-up varies between 8% and 9% in households in which both spouses were eligible for the pension. It also shows that 57% of wives were eligible before their husbands while, vice versa, a higher percentage of husbands, 8%, was. In addition, the third column in both panels show a highly significant difference in take-up rate between treated and control households, regardless of 8 In addition, as a robustness check we also study the three months effect of the pension by setting the end of June as the date in which we observe if one or more spouses was 6 or older. 9 For all variables shown in Table the share of missing values is lower than % in the age range considered. 8

10 whether we consider as treated those households with both spouses being eligible since they are 6 or older or those in which at least one spouse is 6 or older. The panel in the middle of the table shows that, in households in which both spouses are eligible for the pension, poverty rates are smaller than in other households although the difference is only significant for the extreme poverty measure. It also shows that labour supply is significantly lower and consumption is generally higher. Finally, when we look at family arrangements, the relative frequency of adult children living with their parents is lower while, instead, the frequency of male children working and of grandchildren living with their grandparents is significantly higher. Differences in our outcomes of interest between the two groups of households may be driven by other factors than pension eligibility, as we also observe a number of significant differences in household baseline characteristics, such as the relative frequency of urban residents in the bottom panel of the table. This highlights the importance of assessing whether the correlation between being eligible for the pension and our outcomes of interest has a causal interpretation. 3 Conceptual framework Renta Dignidad non-contributory pension in Bolivia consists in a monthly and life-long cash transfer to all individuals who were 6 or older. Since the age condition is the only one an individual needs to fulfill to receive the pension, its recipients can afford greater consumption bundles than before obtaining it, i.e. a pure income effect. The size of the income effect in a household with two elderly spouses depends on the number of pensions observed, that may vary from to 2. In addition, when one pension is observed, the gender of the recipient may vary across households. Finally, the absence of a well-functioning credit market in Bolivia, as well as in other developing countries, may mitigate anticipation effects. Hence, the pension eligibility rule and variation in spouses age across households around the 6 age eligibility cutoff, as Figure 2 shows, offers We discuss further evidence on the difference in households baseline characteristics in section 5.. 9

11 a highly valuable non-experimental exogenous shift in non-labour income to assess changes in household decisions, such as labour supply and consumption. This is important, firstly, to test whether theory models are capable of predicting behavioural responses to the pension in Bolivia and, secondly, to inform policy decisions over the design and implementation of non-contributory pensions. [Figure 2 about HERE] In a decision between leisure and consumption of a normal good, a positive income shock makes the same consumption level affordable by working less, higher consumption with no change in labour supply or all combinations in between these two extreme cases. In addition, labour supply responses may vary by gender. Empirical evidence in developed countries shows that female labour supply is more elastic than for males (Blau and Kahn, 27), with theoretical arguments based on gender roles (Albanesi and Olivetti, 27) or on intrahousehold bargaining between spouses, (Chiappori, 988, 992). In contrast, little evidence is available about gender differences in labour supply elasticities in developing countries, where high labor market informality may induce both females and males to continue working until old age. While the consumption response to an income shock tends to be positive, it may vary substantially by type of goods or services. For example, evidence from developing countries shows that non-durables tend to be elastic to income shocks (Abler, 2; Regmi et al., 2). As for poor households, the pension income shock can be exploited to test for credit constraints by assessing whether consumption of goods such as basic food and medicines increases as soon as first pension payments are received. Living arrangements in a household may also respond to pension income shocks, so as to improve its members living conditions. This holds, in particular, in households in which extended families are observed, that is not infrequent in developing countries. In such households the positive income shock for elderly members may improve living conditions of members not receiving the pension, such as young adults or children. This can be rationalised as beneficiaries altruism towards other family members (Becker, 974) or as an exchange in which they share their pension to obtain, for example, assistance by younger members (Cox, 987).

12 If members of a household not receiving any pension indirectly benefit from the pension received by another one, the pension generates positive spillover effects, that are observed in several policy interventions in developing countries (Angelucci and Di Maro, 25). This can come in the form of in-kind or monetary transfers and the rationale behind is the existence of informal agreements among households engaged in a long-term relationship, i.e. contracts with limited commitment and enforceability (Albarran and Attanasio, 23; Dubois et al., 28). In addition, theory and evidence from Progresa show that spillovers on consumption and investment decisions on non-eligible households are related (Angelucci and De Giorgi, 29; Angelucci et al., 22). Finally, a pension may jointly affect more than one household decision at the same time. The prevalent presumption was that consumption smoothing to a negative income shock is desirable from a welfare viewpoint (Townsend, 994; Morduch, 995). However, Chetty and Looney (26) challenge it theoretically by pointing out that smoothing may hide responses to a shock that may make a household worse-off in the future. This may, for example, occur if children are taken away from school to help with field work. Related empirical evidence from Thai households shows that the presence of kins in a household makes its decisions less dependent on interhousehold transfers thanks to the income kins can generate either through self-production or in the labour market (Kinnan and Townsend, 22). 4 Research design In this section we describe the empirical research design that we use to estimate whether benefits arising from the Renta Dignidad pension income, that increases discontinuously and permanently when individuals turn 6, influence household decisions over consumption, labour supply and family arrangements.

13 4. Regression discontinuity design Let Y be the outcome variable and A an individual s age. A linear regression of Y on a dummy P equal to if the individual obtained the pension and otherwise leads to spurious estimates if individuals unobservable characteristics correlate with whether they received a pension and how long they had received it for. Thanks to a regression discontinuity design (RDD), we can, instead, obtain clean estimates of the pension effect by exploiting the discontinuity at age 6 in the eligibility for the pension to identify the pension effect locally at the age 6 cutoff. Since individuals who are eligible for the pension must apply for it to receive it, receiving the pension is not solely a function of eligibility. By considering the pension as the treatment, comparing outcomes of eligible and non-eligible individuals close to the 6 cutoff under imperfect compliance with the treatment identifies the average effect of assignment into treatment or the intention to treat effect (ITT) at the cutoff. With incomplete take-up, the ITT effect is lower in absolute value than the pension effect has on those who actually receive it (ATT). This is why we estimate the pension effect by instrumenting pension receipt P with a pension eligibility dummy D A equal to if an individual is 6 or older and otherwise by using a fuzzy RDD (Hahn et al., 2; Imbens and Lemieux, 28). Hahn et al. (2) proved the relationship between a treatment effect in an instrumental variables setting and in a fuzzy RDD. Hence, the pension effect can be identified at the cutoff by dividing the jump in the relationship between pension eligibility and the outcome, i.e. the ITT, by the fraction of individuals actually obtaining the pension, as shown in equation (). This is the definition of a Local Average Treatment Effect (LATE), i.e. the ratio between the ITT and the fraction of individuals taking the treatment among all those eligibles or first stage, locally at the cutoff. β = lim ɛ E[Y A = 6 + ɛ] lim ɛ E[Y A = 6 + ɛ] lim ɛ E[P A = 6 + ɛ] lim ɛ E[P A = 6 + ɛ] () Although not all individuals who were eligible for the pension claimed it, the extent of the fuzziness is low since more than 8% of those eligible in the age range 6-8 claimed the pension, as Table shows. 2

14 Receiving a pension may arguably affect decisions taken by an individual, such as labour supply and consumption. However, such decisions may also be influenced by an individual s spouse and may vary depending on whether no spouse, one or both spouses receive a pension. Hence, to study the pension effect on decisions taken at the household level, we modify the research design defined in equation () as follows. Let A F and A M measure age of the female and male spouse in a household and also D Ai = I{A i 6} be dummies equal to if spouse i in the household was 6 or older and, hence, eligible for the pension, and otherwise, with i equal to M for male spoused and F for females. Thanks to rich individual-level data on household members in Bolivia including the day, month and year of birth, we can estimate the difference in an outcome between households obtaining one or more pensions as at least a spouse was 6 or barely older and those that did not as both spouses were barely younger. By adding cross-gender interactions between the age indicator D Ai and age rescaled at 6, i.e. A i 6, we obtain in the equation system in equations (2)-(3) the empirical specification that we use to separately estimate the pension effect by gender. 2 Y hh = β + β P AM + β 2 P AF + β 3 P AM P AF + (2) +f(a F 6, A M 6, D AF, D AM ) + U P i = δ + δ D AM + δ 2 D AF + δ 3 D AM D AF + (3) +f(a F 6, A M 6, D AF, D AM ) + U 2,i,i = F, M, F M Equation (2) estimates the effect that the male spouse obtained the pension (β ), that the female spouse did (β 2 ) and that both spouses did (β + β 2 + β 3 ) by exploiting a RDD with two forcing variables: the age of the female and of the male spouse (Papay et al., 2; Zajonc, 22). We use this design to separately estimate the pension effect for each spouse and for both of them by controlling for the trend of the outcome in the difference in each spouse s age from the 6 cutoff. 2 We can estimate the pension effect by gender thanks to variation in age between spouses, with about 57% of wives obtaining the pension when their husband did not and, vice versa, 8% for husbands, as shown in summary statistics in Table. 3

15 f( ) = β 4 (A M 6) + β 5 (A F 6) + β 6 (A M 6)(A F 6) + (4) +β 7 D AM (A M 6) + β 8 D AF (A F 6) + β 9 D AM (A F 6) + +β D AF (A M 6) + β D AM (A M 6)(A F 6) + +β 2 D AF (A M 6)(A F 6) + β 3 D AF D AM (A M 6) + +β 4 D AF D AM (A F 6) + β 5 D AF D AM (A M 6)(A F 6) In equation (2)-(3) we use the polynomial in spouses age rescaled at 6 and their interactions, defined in equation (4). Parameters β 5 β 8 capture trends in husbands and wives own age and are allowed to differ by whether age is 6 or greater. As for β 9 β 5, they capture gender differences in age trends. The RDD identifying assumption is absence or imperfect sorting by individuals on either side of the age 6 cutoff. In other words, whether an individual is barely younger or older than 6, for example in February 28 when the pension was enacted, is a stochastic shock due to nature. Although the identifying assumption is untestable, date of birth and, hence, age in a small neighbourhood of the 6 cutoff is arguably exogenous. In section 5., we will carefully discuss evidence in support of it in the empirical analysis, by assessing whether individuals age and baseline characteristics are continuous at the 6 cutoff. 3 4 Estimation We estimate equations (2)-(3) by using a local linear regression in spouses age rescaled at the 6 cutoff and Two Stages Least Squares (2SLS). Then, we interpret our estimates as the average effect of the treatment on compliers at the cutoff, i.e local LATE. 4 Since data on the outcomes of interest were collected yearly at the beginning of November and individuals obtained the pension about a month after turning 6, we estimate the pension short-term effect by setting the 3 See Imbens and Lemieux (28) and Lee and Lemieux (2) for additional details about the identifying assumption in a regression discontinuity design. 4 Local linear regressions provide a nonparametric way of consistently estimating the treatment effect in a RDD (Hahn et al., 2; Porter, 23). 4

16 age indicator D Ai = I{A i 6} equal to if spouse i had turned 6 at the end of September and otherwise, as Figure shows. For the sake of robustness, we also estimate the three months effect by setting the age dummy to for individuals turning 6 at the end of June and otherwise. In addition, we drop observations of those individuals turning 6 between June and September, as they have been eligible for the pension for less than three months when outcomes are observed in the household survey in November. 5 Studying the pension effect by gender thanks to a fuzzy RDD with spouses age as running variables, requires choosing the bandwidth around the age 6 cutoff for each of them. If it is too small, estimates are imprecise. If, in contrast, it is too large, they may be biased as the comparison of the outcome value on the right-hand side of the cutoff with the value of the left-hand side is little meaningful. Said differently, the data sample chosen includes observations far away from the age cutoff and, hence, the slope of the polynomial in age rescaled at 6 is not properly accounted for. 6 We select the optimal bandwidth by using the cross-validation based criterion in Papay et al. (2), that consists in four steps. Step : an arbitrary bandwidth value is chosen to regress an outcome on the running variables and on their interaction, separately for subgroups of households in which both spouses were eligible for the pension, in which only one was, and in which nobody was. Step 2: predictions and prediction errors are obtained for all possible pairs of age values for male and female spouses by using estimates obtained in step over a range of age values around the 6 cutoff. Step 3: steps -2 are repeated for different pairs of bandwidth values. Step 4: mean squared prediction errors for each pair of bandwidth values are computed to select as optimal joint bandwidth the one minimizing the mean squared prediction error. 7 In addition, we assess the 5 By dropping individuals becoming eligible for the pension between June and September we can compare households in which at least one spouse obtained the pension for at least 3 months and those receiving no pension at the time of the survey. 6 Bandwidth choice corresponds to choosing the width of the age bins used to compute means of an outcome, that are then plotted to graphically visualise the discontinuity (Lee and Lemieux, 2). 7 We preferred the bandwidth choice criterion proposed in Papay et al. (2) over the one proposed in Zajonc (22) as the latter one is suitable to estimate a treatment effect locally at a discontinuity but only conditional on specific values for the running variables. 5

17 sensitivity of the estimates to using bandwidth values different than the optimal one. In all specifications we use a rectangular kernel, i.e. we estimate a linear regression over a window of age rescaled at the 6 cutoff of width equal to the local bandwidth value on both sides of the cutoff. 8 We also correct standard errors by using the sampling weights in the survey. These are the same when we use data at the individual and at the household level, as the number of households in the primary sampling unit varies from 8 to 35 based on the population density. Finally, we add as baseline characteristics dummies for year, ethnicity, education, residential area, health status and eligibility for a contributory pension to increase the precision of the estimates. 5 Results In this section we firstly show estimates of the pension take-up and of its effect on household decisions in section 5.. In addition, we discuss the validity of the research design in section 5. Finally, we show in Appendix A that they are robust to several changes to our preferred econometric specification and to the definition of key variables. 5. Pension effect Estimates of the short-term pension effect on our outcomes of interest using equation (2) are shown in Table 2, that is divided into five panels by type of outcome. In each regression we report pension effect estimates by gender and when both spouses obtained it, along with the associated p-value of the null hypothesis of no effect. 9 We used optimal age bandwidths to select the data sample used in each regression, following the cross-validation procedure in Papay et al. (2), and report them in the two bottom rows 8 A rectangular kernel has been chosen thanks to its simplicity with respect to more sophisticated ones, since kernel choice tends to have little impact in practice (Imbens and Lemieux, 28). 9 We tested the effect that both spouses received the pension by testing as null hypothesis the linear restriction β + β 2 + β 3 = in equation (2) and computing the p-values of the test. 6

18 in each panel. 2 [Table 2 about HERE] Panel A shows that pension take-up increases discontinuously at 6 by about 6 percentage points, pp hereafter, when only a spouse obtained it and also by about 47 pp when both did. The pension amount received by households with one or more pensions also increases significantly by 2-37 bolivianos with respect to households in which no pension is observed, counterfactual households hereafter. The remaining two columns show estimates of the short-term pension effect on the probability that a household income is below the poverty line and the extreme poverty line, respectively, both measured using a dummy. Although poverty increases by about 5 pp, or by 3% of the mean value of counterfactual households, hereafter shown in parenthesis, when only one pension is received and, in contrast, extreme poverty decreases by up to 2 pp when one or more pensions are observed, these effects are not significant. 2 Panel B in Table 2 shows estimates of the pension effect on household income. The first two columns show that total income and total income per capita increase between 235 and 3 bolivianos (-5%) depending on who obtained the pension although the effect is again not significant. The last two columns show a decrease by 236 bolivianos (245%) in transfers received from other households and, in addition, an increase by 98 bolivianos (35%) in the transfers sent to other households when the male obtained the pension. As for panel C in Table 2, it shows estimates for labour supply. The probability that a wife works decreases by about 27 pp (43%) when her husband obtained the pension and by 36 pp (59%) when both obtained it although the effect is weakly significant. Conversely, the effect on the probability that the husband works is not significant. In addition, the probability that one or more spouses in a household works tends to decrease when one or more pensions are observed in it although the only significant 2 Additional information about the optimal bandwidth calculation is available in section 4. 2 The effect is stronger although still not significant, leading to a decrease in extreme poverty by about 4% relative to counterfactual households not receiving any pension, if it is measured after three and six months since the first pension payment was first made. Results available upon request. 7

19 effect, by about 2 pp (23%) albeit weakly, is observed when both spouses obtained it. Finally, a pension to the male decreases the probability that both spouses work by 32 pp (56%) although weakly. while the negative effect if both spouses receive a pension is not significant. 22,23 As for panel D in Table 2, it shows estimates for consumption. The sign of the effect on total food consumption depends on the beneficiary s gender while the one on non-food consumption is positive. Although point estimates vary between 2% and % of the mean value for ineligible households, they are not significant. In addition, the probability of purchasing medical services increases by almost 32 pp (89%) when a male obtains the pension, although it is weakly significant. Finally, the probability of purchasing medicines increase by about 65 pp (4%) when one or more pensions is observed in a household. 24 Finally, panel E in Table 2 shows estimates of the pension effect on several measures of family arrangements. When only one spouse obtained the pension neither the probability of observing their male or female children, i.e. young adults, working nor that that of observing grandchildren in that household changes significantly. In contrast, in households in which we observe a pension for both spouses the probability that male children work increases by 55 pp (3%), the one that grandchildren live with grandparents by 5 pp (268%) and they are highly significant. [Figure 3 about HERE] In addition, we plotted RDD polynomials of our outcomes of interest to visualise the pension effect. Panel A in Figure 3 shows the pension take-up measured contemporaneously, i.e. one month after turning 6, separately by whether individuals were 6 or older and also by gender. Fitted values of RDD polynomials in age are shown using thick continuous lines and confidence intervals using dashed lines. The figure shows that take-up increases 22 We measure medical services or medicines consumption by using a dummy to test for a qualitative change in their consumption in poor households obtaining one or more pensions as they are services only consumed in the event of illness, whose monetary value follows a skewed distribution peaking at zero. 23 The results are similar when using measures of the intensive margin of household labour supply and are available upon request. 24 The results are similar when using different measures of these household decisions and are available upon request. 8

20 discontinuously at 6 for females and males, from to about 7% and it is highly significant, as the non-overlapping confidence intervals at the cutoff show. These estimates are qualitatively in line with those shown in Table 2 while their precision is not shown as two dimensional plots are little suitable to show changes in an outcome due to simultaneous changes in two running variable across age cutoff, as suggested by Papay et al. (2); Zajonc (22). In addition, Panel A shows a negative although not significant effect on extreme poverty. Panel B to E in the same figure show the same information for our outcomes of interest as follows: income in panel B, labour supply in Panel C, consumption in Panel D and living arrangements in panel E. The only significant pension effect shown in the figure is an increase in the probability that grandchildren live with pensioners. 25 Had we, instead, estimated the pension effect without accounting for the gender of the beneficiary we would have obtained similar results on labour supply decisions, although less precise, and also on consumption, as shown in Appendix B. In contrast, we would have failed to uncover substantive effects on living arrangements involving beneficiaries children and grandchildren. In addition, had we also failed to account for the fact that certain decisions, for example over living arrangements, may be taken jointly by spouses in a household and depend on their bargaining power, we would have mistakenly concluded that the pension has no effect on beneficiaries labour supply, as shown in Appendix C. 5 Research design validity In this section we firstly discuss evidence in support of the identifying assumption and then we present a set of falsification tests. 5. Forcing variables and baseline characteristics Consider an experiment in which subjects are barely younger or older than 6, the treatment consists in receiving a pension and the assignment of subjects to the treated group is random. A RDD identifies the effect of 25 We preferred to show plots of the pension effect in 2 dimensions using a RDD with 2 forcing variables as it gives a qualitative but easy to visualise estimate of the jump at the cutoff. In contrast, 3 dimensional plots show how an outcome varies with values of each running variable although this is less easy to visualise. 9

21 the pension at the age 6 cutoff by mimicking such an experiment in a non-experimental setting. The untestable RDD identifying assumption is that individuals are unable to sort themselves into the treated group, or out of it, by manipulating, for example, their age or date of birth. We offer evidence in support of this assumption by assessing empirically, first, whether the distribution of age by gender is smooth at the age 6 cutoff and, second, whether individuals baseline characteristics are balanced at the cutoff. 26 Since an individual s date of birth is a predetermined characteristic, we expect a smooth distribution of age at 6 for both females and males, in the absence of manipulation. In contrast, an age histogram showing that the height of the bin on the right-hand side of 6 is, for example, higher than for the one on the left-hand side suggests that individuals sorted themselves into the pension. 27 [Figure 4 about HERE] The top panel in Figure 4 shows age histograms, separately by spouse gender in a household, with 9 days as bin width. This ensures that a jump can be detected visually as no bin contains the cutoff value 6. Visual inspection suggests no suspicious jump in the height of the histogram bins at the cutoff, hence supporting the validity of the research design. This result is confirmed by a density-based test of the null hypothesis of no manipulation, i.e. continuous density, at the age 6 cutoff, using the procedure proposed in McCrary (28), as shown by the overlapping confidence interval dashed lines at the cutoff in the bottom panel in Figure 4. [Figure 5 about HERE] In addition, in the absence of manipulation, we expect no difference in individuals baseline characteristics between those who were barely younger than 6 and those barely older, as such characteristics are predetermined. 26 See Lee and Lemieux (2) for a discussion of RDD identifying assumption and validity. 27 See Camacho and Conover (2) for an example of manipulation of a score based on on demographics and income at a cutoff used to allocate individuals to a social program in Colombia. Individuals not truthfully reporting this information and complacent local politicians lead to the observed jump at the cutoff. 2

22 Figure 5 shows fitted values of RDD polynomials in age of each baseline characteristic, estimated separately for individuals 6 or older and for younger ones, as well as by gender. Overall, differences in baseline characteristics at the cutoff are small and not significant, as shown by the overlapping confidence intervals dashed lines. Results are unchanged if we estimate differences in baseline characteristics at the 6 cutoff by regressing each of them on the RDD polynomial in equation (2), as shown by small and non-significant estimates in Table [Table 3 about HERE] 5 Falsification test Finally, we assessed whether 28, the year in which Renta Dignidad was introduced, and 29, are the only ones in which a change in our outcomes of interest at age 6 is observed by estimating the effect of turning 6 in 26 and 27, before the pension was enacted. [Table 4 about HERE] Estimates obtained from equation (2) by using data for the years and defining as treatment group households with spouses turning 6 in 26 and 27 and as control group those with spouses younger than 6 are shown in Table 4. Small and not significant estimates exclude that our main results are driven by confounding factors at 6, such as demographic ones or anticipation effects of the pension, thus supporting the research design validity We also tested the null hypothesis that differences in all baseline characteristics at the age 6 cutoff are jointly zero by estimating a system of seemingly unrelated regressions with as many equations as are baseline characteristics. We obtained them by, firstly, regressing each outcome on age dummies for each spouse equal to if a spouse was 6 or older and on the RDD polynomial in the difference in their age from 6 as seemingly unrelated regressions and, secondly, by testing the null hypothesis that the pension effect on all outcomes is equal to zero. If, for example, β i is the coefficient associated to the female age dummy in equation for outcome i and k is the number of outcomes, the null hypothesis is β = β 2 =... = β k =. Overall, large p-values in Table A.3 in the Appendix suggest that baseline characteristics are jointly balanced when we consider the pension effect for either spouse or for both, thus supporting the research design validity. 29 The results are similar for the 3 months effect of the pension, i.e. when setting the age dummy equal to if an individual obtained the pension at the end of June, as shown in Table A. 2

23 6 Discussion Our results speak to at least three strands of the recent literature in development economics in addition to being highly policy-relevant. Firstly, a non-significant pension effect on poverty, while at first puzzling, is then rationalised by the substantive and highly significant change in a number of household decisions that benefit both pensioners but also, indirectly, their extended family by way, for example, of increased labour supply of young adults, and other households by way of a net decrease in transfers received from them. These results extend the recent literature studying how related decisions of different members of the same extended family or of different households in a networks affect their welfare theoretically (Chetty and Looney, 26; Angelucci et al., 22) and empirically (Angelucci and De Giorgi, 29; Angelucci and Di Maro, 25; Kinnan and Townsend, 22). From a policy viewpoint, this emphasises the importance of anticipating and quantifying potentially unintended policy outcomes at the design stage. Perhaps the most illustrative example is one observed when a pension is received by both spouses. It leads to a decrease in the labour supply by one or more pension beneficiaries by 2pp (23%) that is weakly significant and quantitatively smaller than the indirect and significant increase in the labour supply of their male children, i.e. young adults, by 55 pp (34%). Secondly, finding that the beneficiary s gender matters more for certain household decisions, such as interhousehold transfers, rather than for others, such as purchasing medicines, contributes to empirical evidence on bargaining power in intrahousehold decisions (Attanasio and Lechene, 24; Bobonis, 29; Rangel, 26) by highlighting the importance of assessing the relative contribution of different household members to one or more decisions. From a policy viewpoint, this emphasises the importance of carefully assessing the institutional setting in developing economies such as, for example, prevalent gender roles in households and the implications they may have on individuals and household decisions. Thirdly, finding that the pension decreases interhousehold transfers overall due to a greater decrease in transfers received than the increase in transfers sent when a male obtains the pension adds evidence to the 22