94662 GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION: EVIDENCE FROM AN UNINTENTIONAL EXPERIMENT Luis H. B. Braido, Pedro Olinto, and Helena Perrone∗ Abstract—We use data from a Brazilian social program to investigate the program required compliance with vaccination schedules and existence of gender bias in intrahousehold allocations of resources. The program makes cash transfers to mothers and pregnant women in poor regular visits to public centers for prenatal care, child-growth households. Bureaucratic mistakes, beyond the control of the applicants, monitoring, and health education classes. The transfers were have inadvertently excluded many households that had applied and were always made to the highest-ranking woman in the household, accepted to the program. This unintentional natural experiment is used to identify the impact of an exogenous variation in female nonlabor income usually the mother of all the children in the family. Payments over household consumption. We find that program participation led to an were conducted by a large federal bank, Caixa Econômica increase in food expenditure, but this effect is not due to women being the Federal (CEF), which issued a personalized ATM card to each benefit recipients. beneficiary woman. These electronic cards could be used at CEF branches and network locations, as well as in authorized I. Introduction retail stores. The amount transferred averaged approximately 8% of C ONDITIONAL cash transfer (CCT) programs pervade virtually every country in Latin America, collectively benefiting more than 70 million people throughout the total household expenditure. Thus, participation in the pro- gram represented a significant exogenous variation in female region. Following this lead, CCT programs have also been nonlabor income. Beneficiary families likely believed the launched throughout the world, including Bangladesh, Cam- increase in income would last for a long period because as bodia, Indonesia, Malawi, Morocco, Pakistan, South Africa, their children grew older, they would become eligible to the and Turkey. More recently, Washington DC, and New York follow-up CCT program, the Bolsa Escola (BE). The BE pro- City have launched pilot programs that use CCTs to fos- gram targeted families with children between ages 7 and 14 ter household investments in children’s schooling. Overall, years of age and was conditional on school attendance.1 Note, these programs are widely perceived as being effective in however, that even a short-term transitory shock on female reducing poverty in both the short and long runs (see Lindert, income could affect intrahousehold consumption decisions, Skoufias, & Shapiro, 2006; Fiszbein et al., 2009). since it is reasonable to believe that poor families in Brazil Almost all CCT programs share the common feature that are credit constrained. the cash transfer is made to a woman. This policy design is motivated by a growing belief that women exhibit expen- A. Accidental Exclusions diture patterns that are more pro-child and pro-family than men do. This is also supported by a solid theoretical literature The uniqueness of the data lies in the fact that a group of arguing that family expenditures may depend on factors such eligible households, which had applied and were accepted as the intrahousehold distribution of income. to the program, were randomly and unintentionally excluded This paper empirically assesses whether poor women from it. The exclusion of eligible households occurred due empowered by CCTs do indeed divert household resources to three independent reasons, all of them beyond the control toward goods typically considered as pro-child and pro- of the applicants and unrelated to households’ unobserved family. In pursuing this goal, we use a special data set characteristics. collected in 2002 to evaluate the impact of Bolsa Alimen- First, some files containing household-identifying infor- tação (BA), one of the first CCT programs implemented in mation were lost during electronic transmission to CEF, Brazil. The empirical strategy relies on the fact that many eli- the bank responsible for the payments. These exclusions gible households that had applied and were accepted to the were due to local network problems and staff mistakes. program were randomly and unintentionally excluded from it. For each municipality, they are independent of household The evaluation focuses on poor areas of the northeast region characteristics. of the country. The second source of exclusion stems from the software The BA program aimed at reducing infant mortality and used by municipal authorities in charge of program registra- nutritional deficiencies among children from very poor fam- tion: although this software was adapted to the Portuguese ilies. It started in 2001 and consisted of cash transfers to language, the software used in the federal capital by CEF pregnant women and mothers of children under age 7. The for issuing the beneficiary identification number and pro- cessing payments was not. In fact, writing names in capital Received for publication July 23, 2008. Revision accepted for publication letters and without accent marks is a convention in the Brazil- September 29, 2010. ian banking system. Thus, because the CEF software was ∗ Braido: Getulio Vargas Foundation, Graduate School of Economics; Olinto: World Bank; Perrone: Universitat Pompeu Fabra and Barcelona GSE. We are indebted to Marco Bonomo, Sergio Firpo, and anonymous referees 1 These two programs, the BA and BE, were unified in 2003 and are now for insightful remarks. called Bolsa Família. The Review of Economics and Statistics, May 2012, 94(2): 552–565 © 2012 by the President and Fellows of Harvard College and the Massachusetts Institute of Technology GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 553 not able to read special characters (such as ç, ã, é, and ô), since they live in the same municipality and their choices are households in which at least one member had any special observed in the same period. This represents an important character in the name did not receive an identification number advantage over quasi-experimental before-and-after analy- and were not included in the program roster for some time. ses, as well as over traditional randomizations in which the These characters are commonly found in Brazilian names program implementation is delayed in some randomly chosen (such as João, José, Ângela, Andréa, Tânia, and Mônica) locations. and surnames (such as Aragão, Gonçalves, Magalhães, Men- The survey design and some descriptive statistics are pre- donça, and Simões). Furthermore, these names and surnames sented in section III. Next, in section IV, we discuss the are homogeneously distributed across the population and are econometric conditions under which OLS regressions consis- not linked to specific ethnic, gender, age, or income groups. tently estimate the average treatment effect of the BA program Hence, this second exclusion criterion is also exogenous, on the subpopulation of accidentally excluded households in conditional on the number of members in the household.2 those areas. We also show that when compared to nonbenefi- The third source of exclusion is related to misspelling prob- ciaries, BA participants spend proportionally less on utilities lems. Many households that applied and were eligible for the and more on vegetables and fruits. We do not find statisti- BA program had school-aged children and were also enrolled cally significant impacts of the program on expenditure shares in the BE program. The federal bank responsible for the trans- of dairy, vices (alcohol, tobacco, and gambling), and family fers of both programs, CEF, decided that it would issue a goods (such as health, schooling, clothes, personal hygiene, single identification number for each family, hoping that in and house cleaning products). the future, this would become a single social security number The relative increase in the consumption share of vegeta- to be used by all federal programs. Therefore, CEF blocked bles and fruits could result from three sources: (a) a female the registration of households whose data arriving from the empowerment effect due to the increase in the fraction of BA registration showed any inconsistency with the informa- household income accruing to the beneficiary women, (b) tion recorded during the earlier BE registration. For instance, an income effect resulting from the program transfer (and if during the BA registration a household member’s name not fully captured by the budget share econometric represen- was spelled differently from what had been recorded in the tation), and (c) a health-monitoring effect resulting from the BE registration, the entry in the BA program was frozen until fact that BA participants committed to regular visits to public this inconsistency was clarified. Since the electronic records health centers. in both programs were filled out by staff members, this third In section V, we present a model of household behavior that type of exclusion is related to administrative problems beyond illustrates how a change in income shares could affect house- the control of the applicants. It is thus exogenous, conditional hold consumption choices. Next, in section VI, we explore on the number of household members, the municipality where the family structure of sampled households to disentangle registration was conducted, and previous enrollment status in potential female empowerment effects from the other two the BE program. possible effects of the program. We measure the impact of Households and local authorities were aware that regis- BA participation on expenditure patterns in two groups: fam- tration did not guarantee participation. Eligibility needed ilies with adults of different genders and families with no to be confirmed by the program staff, who predicted per male adult living in the household, typically composed of capita income for each family using national surveys and the single mothers and their children. We refer to these groups as household information collected during the registration pro- mixed-gender and female households, respectively. Gender- cess. Moreover, there was great uncertainty about the number specific empowerment effects should be present only in the of beneficiaries since the program budget allocated to each first category of households. municipality was not announced in advance by the federal This difference-in-differences (diff-in-diff) strategy is government. It is thus reasonable to assume that households valid under the identifying assumption that the BA program interpreted these accidental exclusions as being caused by did not affect the fraction of female households. Naturally the lack of eligibility and did not act on it. (The possibility of pro- BA program could have affected the family structure through gram reinclusions caused by households’ actions is addressed changes in divorce rates. We show, however, that the condi- in section IVB.) tional correlation between BA participation and the family structure is not statistically significant. Fortunately, it seems B. The Analysis that those potential changes have not systematically occurred during the first six months of the program (when the data were We start our analysis by reviewing the related literature in recorded). section II. We point out that our accidentally excluded and According to our results, the change in income shares of matched beneficiary households face the same relative prices, men and women did not affect household consumption pat- terns. The BA program did affect household expenditures on 2 It will be necessary to control for the household population in the econo- utilities, vegetables, and fruits, but the average impact is sta- metric estimations because larger households are more likely to have at least one name with a special character and household size might be correlated tistically equal across female and mixed-gender households. with the outcome variables of interest. Similar results are obtained when we compare two groups 554 THE REVIEW OF ECONOMICS AND STATISTICS of mixed-gender households—one in which the beneficiary household characteristics. The demand curves of beneficiary women have another source of personal income different households shifted up for children-assignable goods (such as from the BA transfer and another in which they do not. These clothes and toys) and shifted down for men’s clothing and results and additional robustness exercises appear in section tobacco. VI. A brief conclusion is presented in section VII. A similar quasi-experiment took place in Australia dur- ing the 1990s and was analyzed by Bradbury (2004). During II. Literature Review the decade, a series of public reforms resulted in signifi- cant changes in transfer payments to married-couple families. A large literature is dedicated to the analysis of gender dif- The previous system, in which payments were almost solely ferences in household expenditure decisions. We are aware received by men, evolved to one in which more than half of of the following works that, similar to ours, use natural exper- the public transfers were directed to women. Bradbury esti- iments or quasi-experiments to study the effect of women’s mates Engel curves for 23 different commodities in order to empowerment policies on household allocations. assess the effect of the changes on the intrahousehold distribu- Attanasio and Lechene (2002) examine the effects of par- tion of income on expenditure patterns. The analysis focuses ticipation in a cash transfer program, PROGRESA, which on the Australian Household Expenditure Survey, which is started in 1998 in rural Mexico. The transfer is conditional conducted every five years. Three surveys are used: 1988– on school enrollment, and the mother is always the recipi- 1989 (before any policy change), 1993–1994 (when part of ent. As part of the program design, a number of randomly the reform had already taken place), and 1998–1999 (after selected villages did not start the program for a year and a the last policy change). Contradicting the previous literature, half; these communities form the control group. The authors results indicate a very small effect of the income shift on consider eight types of nondurable expenditures: food, alco- household expenditure allocation. Furthermore, the few sig- hol and tobacco, transportation, services, and clothing for nificant changes were frequently in the wrong direction—that women, men, girls, and boys. The empirical results indicate is, increases in the expenditure share of goods such as tobacco a positive impact of female income on girls’ and boys’ cloth- and alcohol and decreases in the expenditure share of food ing and a negative impact on alcohol. They also show that consumed at home. self-reported decision-making power varies across recipients, Our work has two important advantage over these exer- nonrecipients, and future recipients in the control villages. cises. Unlike those in Attanasio and Lechene (2002), our In the quasi-experimental literature, Lundberg, Pollak, treated and nontreated households live in the same municipal- and Wales (1997) use the 1979 reform in the U.K. family ity. Moreover, different from the before-and-after exercises allowance policy. This reform replaced a tax allowance pro- in Lundberg et al. (1997), Ward-Batts (2008), and Bradbury gram with a nontaxable payment made directly to mothers. (2004), we analyze treated and nontreated households dur- To the extent that tax allowances benefited mainly the fathers, ing the same period. These two features allow us to avoid this reform transferred a substantial amount of income to the concerns about eventual changes in relative prices across mothers. The data include the average consumption of clothes municipalities or over time. It is worth stressing that as in for different family categories (defined according to income Attanasio and Lechene (2002) and unlike the other sources, and number of children). They use the period 1973 to 1976 to our analysis refers to very poor families that spend most of represent the consumption regime before the policy change their income on basic goods. and the period 1980 to 1990 to represent the regime after the reform. They find that in beneficiary households, expendi- tures increased for children’s and women’s clothing relative A. Additional Related Works to men’s. This analysis was limited to clothing expenditures and did not take into account eventual changes in relative In a related strand of the literature, Browning et al. (1994) prices (as well as other possible changes in the economic explicitly model household decisions under the assumption environment) over the years studied. In fact, by analyzing that the interactions among household members with dif- this tax reform, Hotchkiss (2005) found similar changes in ferent preferences lead to Pareto-efficient outcomes. The expenditure patterns of families with no children and raised structural parameters of the model are estimated using Cana- alternative explanations for Lundberg et al.’s results. dian data on couples with no children. The main goal of the Ward-Batts (2008) uses this same data source disaggre- analysis is to investigate how final outcomes depend on the gated to the household level to test the effect of the 1979 income each member brings into the household. They also reform in the U.K. family allowance policy on household compare expenditure behavior in single-person households expenditure shares. The periods 1973 to 1976 and 1980 to with that of couples. They find that older and higher-income 1983 represent the regimes before and after the reform. She partners are able to divert a higher share of total house- uses data on households consisting of one man, his wife, hold expenditure toward their own consumption. Moreover, and between one and three children under 18 years of age. holding age differences and relative income shares constant, A demand system is estimated using data on consumption women are able to divert more income toward their own expenditures, a price index for each category of goods, and consumption in wealthier households. GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 555 There are also many relevant reduced-form studies that The transfers were conditional on women committing to a estimate the effect of women’s leadership on household- “charter of responsibilities” that required compliance with choice variables. Handa (1996) presents evidence that female- vaccination schedules and regular visits to health centers headed households in Jamaica dedicate a greater budget for prenatal care and child growth monitoring. Beneficiary share to children’s clothes, health, and food goods, while women were also required to attend to health, nutrition, and male-headed households spend more on alcohol and tobacco. child care classes. Eligibility for the BA expired when chil- Thomas (1990) uses a household survey from Brazil (Estudo dren turned 7 years old. Poor families with children between Nacional de Despesa Familiar) to show that nonlabor income 7 and 14 years of age would then become eligible for the in the hands of women has a greater positive effect on chil- Bolsa Escola (BE) program, which ensured continued cash dren’s anthropometrics than nonlabor income in the hands transfer of the same amount as the BA program but con- of men. Thomas (1994), using data from the United States, ditional on school attendance. As in the BA program, the Brazil, and Ghana, shows that mothers’ education level tends transfer recipient was also the highest-ranking woman in the to have a stronger effect on girls’ height (relative to boys), household. while fathers’ education has a larger impact on boys’ height. The special feature of the data is the existence of a control Thomas, Contreras, and Frankenberg (2002) use data from group formed by households that had applied for the program the Indonesia Family Survey to study how child morbidities and were eligible to benefit from it but were unintention- (such as diarrhea, cough, and fever) are affected by the relative ally excluded. As explained before, three types of accidental value of assets that wives and husbands bring to the marriage exclusion were detected: (a) some household files were lost (an indicator of economic independence). The results suggest during the electronic transmission from the local registra- that gender matters in Java and Sumatra, although not in the tion offices to CEF, the national bank responsible for issuing other Indonesian regions considered. identification numbers and transferring the payments; (b) the Although important as an initial assessment of intra- computer program used by CEF was not adapted for the Por- household allocation decisions, this reduced-form evidence tuguese language and then excluded households in which one is potentially biased by the endogeneity of the leadership or more members had special characters in their names; and variables (for example, headship, income share, education, (c) some households previously enrolled in the BE program and asset ownership). Unobserved characteristics of the fam- were excluded from the BA program because CEF blocked ily could be correlated with both leadership and household the registration of those whose data coming from the BA reg- members’ willingness to divert resources toward children- istration showed any inconsistency with the data previously assignable goods.3 recorded for the BE registration. For each municipality, the first type of exclusion is com- pletely exogenous. The second type of random exclusion, III. Data however, is only conditionally exogenous. That is, it is exoge- We use data from a survey conducted by the International nous conditional on the number of household members, since Food Policy Research Institute (IFPRI). This research was larger households were more likely to have at least one name contracted by the Brazilian Ministry of Health in order to with a special character. Likewise, the third type of exclusion evaluate the impact of the BA program on several nutritional is exogenous conditional on the number of household mem- and health outcomes. bers, the municipality where registration was conducted, and Initiated in 2001, the BA program consisted of cash trans- previous enrollment status in the BE program. fers to low-income families with pregnant women or mothers Accidentally excluded households were aware that regis- of children under 7 years of age. Households were eligible tration did not guarantee participation. Municipal authorities if their estimated monthly per capita income was below 90 were asked to register considerably more than their specified BRL (equivalent to about 37.50 U.S. dollars in 2001 values). quotas with the expectation that some registered households The highest-ranking woman in the household, typically the would later be found to be ineligible. The precise number of mother of all the children, was the sole recipient of the cash beneficiaries would also depend on the program budget allo- transfer. The values transferred were 15, 30, or 45 BRL per cated to each municipality. In addition, households were not month (something between 6.25 and 18.75 U.S. dollars). The able to communicate directly with CEF. Therefore, it is plau- exact value depended on the number of qualifying children sible that households typically did not react when excluded, in the household. believing their exclusion was due to lack of eligibility. When the survey interviews were conducted, 19 (out of 282) households in the accidentally excluded group had 3 Two other classes of related works should be mentioned. A number been reincluded in the program and reported receiving the of authors have studied intrahousehold allocation of resources in families consisting of parents and their children or grandchildren; see Costa (1997), BA transfer. Similarly, 44 (out of 717) households in the Pezzin and Schone (1997), Duflo (2000, 2003), and Edmonds, Mammen, matched beneficiary group reported not receiving transfers. and Miller (2005). There is also an important policy debate on whether social These reinclusions and late exclusions were conducted by programs targeted to children should make monetary or in-kind transfers. See Bingley and Walker (2009) for quasi-experimental evidence from the the program staff, who did not communicate directly with United Kingdom. the households. They are likely independent of household 556 THE REVIEW OF ECONOMICS AND STATISTICS characteristics and are treated as such in most of the paper. It is important to stress that we do not have access to We address the potential endogeneity problems associated the presurvey data collected during registration and used by with these late inclusions and exclusions in section IVB. IFPRI for matching beneficiaries to the accidentally excluded households. However, the survey data used in this paper con- tain more information about household characteristics than A. Survey Design what was available for the sampling design. Using the survey Excluded households were found in 67 Brazilian munic- data, we test whether the socioeconomic characteristics that ipalities. Two criteria guided the selection of the munici- are unlikely to have been affected by the program are in fact palities in the evaluation study. First, municipalities had to balanced across excluded and matched households. In table 1, be located in the northeastern region of the country. About we regress the BA participation dummy on different informa- 60% of the beneficiaries resided in this poor region of Brazil. tion about the education of the eligible woman, on a dummy Second, for cost-saving reasons, only municipalities partici- variable describing whether the household is headed by a pating in the program for the previous six months and with woman, and on the number of household members of differ- at least forty excluded families were surveyed. ent gender and age ranges. All regressions are conditional on In April 2002, when the survey team went to the field, the number of household members, the municipality where four municipalities fit these two criteria: Teotônio Villela, in registration was conducted, and previous enrollment status the state of Alagoas; Mossoró, in the state of Rio Grande in the BE program, since accidental exclusions are supposed do Norte; and Itabuna and Teixeira de Freitas in the state of to be random only after conditioning on these variables (see Bahia. All excluded households were surveyed, and a sam- section IV for a formal argument on this linear conditioning ple of matching beneficiaries was selected from the roster of approach). receiving beneficiaries. The data collected during the regis- The reported educational characteristics of the eligible tration process were used to match two beneficiary families to woman, the fraction of female-headed families, and the each excluded household. The matching criteria required that household gender-age structure are conditionally uncorre- participants and nonparticipants exhibited residence in the lated to BA participation. Since these variables were not same municipality, identical gender for each eligible child, used in the matching algorithm, our findings strongly support and similar socioeconomic characteristics. After using the the baseline assumption that exclusions were conditionally first two criteria to determine a preliminary pool of matched exogenous.5 beneficiaries, the following variables were used to match in terms of preprogram socioeconomic characteristics: per B. Summary Statistics capita self-reported monthly income, number of household members, and per capita monthly expenditures on rent, water, The database contains detailed information on household electricity, and gas. Using principal components analysis, expenditures. The survey questionnaire uses modules 1 and these variables were reduced to a single factor. The matching 5 to collect information on nonfood items and module 6 for algorithm consisted on finding, in a random order, optimal food items.6 We aggregate these items in nine expenditure matches based on the weighted sum of the squared dif- groups: ference in the socioeconomic scores and the squared age difference of the eligible women. This technique is known • General services: Water, telephone, gas, and electricity as nearest-neighbor matching based on Euclidean distances (module 1, all items), and gasoline, auto services, taxes, (see Rosenbaum & Rubin, 1985). In a technical report, pensions, lawyers, home insurance, mobile phone, wed- IFPRI (2003) describes the matching algorithm, question- dings, donations, and funerals (module 5, items gr1, gr5, naire design, and instruction for interviewers, among other gr9, gr11, gr13, gr15–gr18) implementation details. • Family expenses: Health, schooling, clothes, personal The use of household characteristics to construct a matched hygiene, house cleaning products, home maintenance, sample was first proposed by Donald Rubin in a series furniture, and maids (module 5, items gr2–gr4, gr6–gr7, of studies summarized in Rubin (2006). This methodology gr12, gr14, gr19–gr21) consists of a systematic approach to sampling control and • Vices: Tobacco and gambling (module 5, items gr8 and treatment groups in order to increase the overlap in the distri- gr10) and alcoholic beverages (module 6, questions 75 bution of covariates of both groups. Regression analysis can and 76) be then conducted in matched samples. Rubin (1979) uses • Grains: Module 6, questions 1–12 Monte Carlo simulation to show how regression adjustments • Vegetables: Module 6, questions 13–27 coupled with a matched sampling procedure can improve estimation results in small samples.4 5 Morris et al. (2004) use a different set of questions from this same ques- tionnaire and report that the originally excluded and the matched households are similar in terms of the type of floor material of their homes, access to 4 See also Heckman, Ichimura, and Todd (1998), Abadie and Imbens water supply through a public network, and access to a telephone. (2006, 2008), and Imbens and Wooldridge (2009) for a discussion on the 6 The data also display imputed values for rent. However, we fear this asymptotic properties of matching estimators. variable is poorly measured and decided not to use it. GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 557 Table 1.—BA Participation and Household Characteristics BA Dummy (1) (2) (3) (4) (5) (6) Beneficiary educational dummies Reads well −0.030 (0.035) Reads poorly −0.037 (0.036) Writes well −0.012 (0.034) Writes poorly 0.003 (0.036) Understands basic math −0.017 (0.034) Poorly understands basic math −0.024 (0.036) Is currently registered at school 0.034 (0.050) Had been registered at school 0.002 (0.036) Female headship dummy −0.009 (0.044) Household composition Number of males 0–6 years old 0.028 (0.057) Number of males 7–14 years old −0.019 (0.059) Number of males 15–18 years old −0.086 (0.068) Number of males 19–60 years old −0.061 (0.057) Number of males older than 61 years dropped Number of females 0–6 years old 0.052 (0.058) Number of females 7–14 years old −0.011 (0.058) Number of females 15–18 years old 0.016 (0.071) Number of females 19–60 years old 0.022 (0.067) Number of females older than 61 years −0.022 (0.093) Number of household members −0.003 −0.001 −0.002 −0.001 −0.001 0.007 (0.008) (0.008) (0.008) (0.008) (0.008) (0.056) BE dummy −0.230∗∗∗ −0.229∗∗∗ −0.229∗∗∗ −0.229∗∗∗ −0.228∗∗∗ −0.209∗∗∗ (0.034) (0.034) (0.034) (0.034) (0.034) (0.037) Municipality dummies Yes Yes Yes Yes Yes Yes Number of observations 1,006 1,006 1,006 1,006 1,006 1,006 OLS regressions with constant; dependent variable in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. • Fruits: Module 6, questions 28–40 The fractions of zero consumption are reasonably small for • Dairy: Module 6, questions 41–50 the remaining eight expenditure categories. • Meat: Module 6, questions 51–63 Table 3 shows the sample means of each variable for partic- • Oils, spices, and soft drinks: Module 6, questions 64–74 ipant and excluded households. It is divided into two blocks and 77–81 according to BE enrollment status. Households enrolled in the BE program must have at least one school-aged child. On The household monthly expenditure, defined as the sum average, they spend more and have a slightly larger number of all those expenditures, is used to construct the expenditure of members. shares. Other variables, such as the number of household members, a dummy variable describing whether the house- IV. Average Treatment Effect hold was already enrolled in the BE program, and dummies for each municipality, are also used in our analysis. Their Consider the standard switching regression framework to summary statistics are presented in table 2. The average frac- model unobserved counterfactuals. Let Yi be an arbitrary ran- tion spent in each group ranges from 1.6% for vices to 20.1% dom outcome associated with household i. If household i is for grains. (It is probably worth mentioning that grains such enrolled in the BA program, its outcome Yi is assumed to be as rice and beans are very common in the Brazilian diet.) About 52% of households report no consumption of vices. Y1,i = μ1 + ε1,i . (1) 558 THE REVIEW OF ECONOMICS AND STATISTICS Table 2.—Summary Statistics Mean S.D. Minimum Maximum Frequency of Zeros Expenditure shares General services (utilities, durables, etc.) 16.0% 8.9% 0.0% 54.0% 0.7% Family expenses (clothes, education, health, etc.) 17.0% 10.7% 0.0% 89.7% 0.3% Vices (alcohol, tobacco, and gambling) 1.6% 3.2% 0.0% 32.8% 52.2% Grains 20.1% 8.4% 0.0% 66.4% 0.1% Vegetables 5.2% 3.9% 0.0% 28.9% 4.6% Fruits 4.7% 4.1% 0.0% 33.7% 12.4% Dairy 7.4% 5.6% 0.0% 50.6% 5.3% Meat 18.6% 8.6% 0.0% 47.7% 1.0% Oil, spices, and soft drinks 9.4% 4.8% 0.0% 63.9% 0.4% Total Monthly Expenditure (BRL) 373.29 166.61 29.00 1, 116.72 0.0% Number of household members 5.5 2.2 2 16 0.0% Participation in the BE program 34.6% 47.6% 0 1 – Municipality dummies Itabuna 17.1% 37.7% 0 1 – Mossoró 26.7% 44.3% 0 1 – Teixeira de Freitas 42.8% 49.5% 0 1 – Teotônio Villela 13.3% 34.0% 0 1 – Number of observations: 1,006 households. Table 3.—Summary Statistics in Subsamples Households Not Enrolled in the BE Program Households Previously Enrolled in the BE Program Treated Group (BA = 1) Control Group (BA = 0) Treated Group (BA = 1) Control Group (BA = 0) Mean S.D. Mean S.D. Mean S.D. Mean S.D. Expenditure shares General services 15.7% 8.3% 17.0% 9.3% 15.0% 8.8% 17.0% 10.3% Family expenses 17.3% 11.0% 17.9% 11.3% 15.7% 9.5% 16.8% 10.3% Vices 1.5% 3.3% 2.1% 3.9% 1.5% 2.9% 1.5% 2.8% Grains 19.1% 8.0% 18.6% 8.7% 22.3% 7.6% 22.1% 9.1% Vegetables 5.6% 4.2% 4.8% 3.6% 5.1% 3.2% 4.5% 3.7% Fruits 5.1% 4.1% 4.2% 3.9% 4.9% 4.9% 4.0% 3.3% Dairy 7.7% 5.6% 8.0% 6.7% 7.2% 4.9% 5.9% 5.0% Meat 18.9% 8.6% 17.9% 8.7% 18.6% 8.1% 18.0% 8.9% Oil, spices, and soft drinks 9.1% 4.8% 9.4% 4.6% 9.6% 4.4% 10.1% 5.1% Total monthly expenditure (BRL) 371.72 162.70 363.38 169.60 384.21 149.08 374.66 194.24 Number of household members 4.9 1.9 5.0 2.2 6.6 2.2 6.5 2.1 Number of observations 509 149 188 160 If the family is not enrolled in the BA program, this same If the program assignment were random, one would have outcome is assumed to be E (ui | BAi ) = 0, and the OLS method would consistently estimate the parameters of equation (5). However, since BA Y0,i = μ0 + ε0,i . (2) participation was accidental, identification cannot be taken for granted. The terms μ1 and μ0 are constant parameters, and ε1,i and ε0,i There were three sources of accidental exclusions: data are random variables with 0 expected value. loss during electronic transmission, special characters in the In this setting, the average impact of the BA program on name of some household members, and inconsistency with Yi is previously recorded data for households already registered in the BE program. The first type of exclusion is independent α = E (Y1,i − Y0,i ) = μ1 − μ0 . (3) of household characteristics for each municipality. The sec- ond type of exclusion is affected by the number of members Let BAi be a dummy variable that equals 1 for beneficiary in the household: the larger the household, the more likely households. Household i’s outcome can be expressed as some family member will have a name with some special Yi = BAi Y1,i + (1 − BAi )Y0,i , (4) character. Similarly, the third type of exclusion is exogenous only once we control for the number of household members, and, hence, the municipality where registration was conducted, and the household enrollment status in the BE program. Yi = μ0 + αBAi + ui , (5) Let Xi be a k -dimensional real-valued random vector con- taining different characteristics of household i, including where ui = ε0,i + BAi (ε1,i − ε0,i ). (in particular) variables describing the number of household GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 559 Table 4.—BA Effect on Log Expenditures the BA dummy are 5.3% (significant at the 10% level) for the Ln Total Ln Food Ln Nonfood log total expenditure regression, 9.9% (significant at the 1% Expenditure Expenditure Expenditure level) for the log food expenditure regression, and 0.6% (not BA dummy 0.053∗ 0.099∗∗∗ 0.006 statistically significant at the 10% level) for the log nonfood (0.031) (0.033) (0.049) expenditure regression. The average BA transfer amounted Number of household members, BE, and municipality to about 8% of the household average expenditure. Thus, the dummies Yes Yes Yes increase of 5.3 log points in total expenditure is meaningful. Household characteristics Yes Yes Yes The higher total consumption in beneficiary households is Number of observations 1,006 1,005 1,006 almost entirely due to higher food consumption. OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 1%, ∗∗ 5%, ∗∗∗ 1%. We then perform disaggregated regressions in which the dependent variable is the household expenditure in each good members, municipality, and participation in the BE program. category. These results appear in table 5. The estimated coeffi- According to the description of the accidental exclusions, the cients associated with the BA dummy are positive for all good BA dummy is exogenous conditional on Xi . Therefore, one categories except general services, family goods, and vices. should have They are statistically different from zero for four good cate- gories: grains, vegetables, fruits, and meats. This means that E ui | Xi , BAi = E ui | Xi = g(Xi ). (6) in absolute values, BA participants consume more of these three types of goods than the families that were accidentally Existence and almost everywhere uniqueness of the func- excluded from the program. tion g is well established in probability theory. We assume Next, the regressions in table 6 use the share of total this function to be linear such that expenditure diverted to each good category as the dependent variable. On average, BA participants spent proportionally g(Xi ) = Xi · θ (7) more on vegetables and fruits and less on general services for a given parameter vector θ ∈ Rk . (such as utilities) when compared to nonparticipants. Equations (5) to (7) lead us to the following regression Later, in section VI, we focus on understanding which model: part of these changes in expenditure shares is due to a female empowerment effect as opposed to income and health- Yi = μ0 + αBAi + Xi · θ + ui , (8) monitoring effects. (Recall that BA beneficiaries experienced an increase in their income and were required to attend health where ui = ui − g(Xi ). education classes, where they probably learned about the The error term ui is orthogonal to (Xi , BAi ). This fol- importance of a healthy diet.) lows from the fact that BA participation is conditionally exogenous, after one controls for the number of household members, the municipality of the household, and the previ- B. Late Inclusions and Exclusions ous status in the BE program. These variables must always be There is an important issue to address before proceeding. included in Xi in order to guarantee consistency of the OLS method. However, following Rubin (1979), we also include in The accidental exclusions described earlier were not immedi- Xi all household characteristics available in the database (see ately detected by program managers. However, once an error the list of these characteristics in the first column of table 1). was detected and fixed by CEF bank, the household was rein- Although the random nature of the program exclusions would cluded in the program. When the survey team went to the field balance covariates across large treated and control groups, to conduct the interviews, they found 19 (of 282) households the small sample results can be improved by surveying a in the accidentally excluded group that had been reincluded matched sample and implementing regression adjustments. in the program and reported receiving the BA transfer. They In Rubin (1979), the regression adjustments account for the also found 44 (of 717) households in the original matched fact that matched households are similar but not identical beneficiary group that reported not receiving the transfers. since the matching index algorithm reduces, but does not Since these late inclusions and exclusions were conducted eliminate, differences across multiple characteristics. In our by CEF staff without any influence from the municipality case, the regression adjustments also control for household authorities and households themselves, they are likely to be characteristics that became available only after the data were conditionally independent of household observed and unob- collected. served characteristics. Nevertheless, to address this issue, we estimate equation (8) using instrumental variables.7 A. OLS Results 7 IV estimation is consistent for the average treatment effect (ATE) under We first analyze the average impact of the BA program over the assumption of homogeneous impacts across all households in our pop- three alternative dependent variables: log total expenditure, ulation of interest. Under weaker assumptions, IV methods are consistent for the local average treatment effect (LATE), which is the ATE for the log food expenditure, and log nonfood expenditure. These subpopulation of compliers, that is, those who comply with the accidental regressions appear in table 4. The estimated coefficients for experiment (see Angrist, Imbens, & Rubin, 1996). 560 THE REVIEW OF ECONOMICS AND STATISTICS Table 5.—BA Effect on Expenditure Levels Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA dummy −3.920 −2.480 −1.541 4.305∗ 3.286∗∗∗ 2.765∗∗ 1.284 6.302∗∗ 0.201 (3.048) (4.646) (1.458) (2.348) (1.030) (1.273) (1.514) (3.137) (1.201) Number of household BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Table 6.—BA Effect on Expenditure Shares Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA dummy −0.012∗ −0.007 −0.003 0.003 0.008∗∗∗ 0.008∗∗∗ 0.001 0.007 −0.004 (0.006) (0.007) (0.002) (0.006) (0.003) (0.003) (0.004) (0.006) (0.003) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Table 7.—BA Effect on Log Expenditures: IV Method mechanisms (see Manser & Brown, 1980; McElroy & Hor- Ln Total Ln Food Ln Nonfood ney, 1981; Lundberg & Pollak, 1993; and Chen & Woolley, Expenditure Expenditure Expenditure 2001). Given the illustrative purpose of this section, we focus BA dummy 0.072∗ 0.108∗∗∗ 0.031 our discussion on the collective approach, which assumes (0.038) (0.039) (0.058) that household choices are Pareto efficient regardless of the Number of household members, BE, and municipality dummies Yes Yes Yes specific details about the underlying bargaining process (see Household characteristics Yes Yes Yes Chiappori, 1988; Browning, 1992; Browning, Chiappori, & Number of observations 998 997 998 Lechene, 2006). IV regressions with constant; dependent variables in first row; controls in first column. Instrumented variable: BA participation dummy; instrumental variable: BA initial status Dummy. Robust standard errors Consider a household composed of one female and one in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. male, respectively indexed by f and m, who consume qf ∈ RL + and qm ∈ RL + units of L > 1 different commodities. Define the household utility function as: We use the BA status when the program started (before V (q, p, w, d ) = max λ(p, w, d )uf (qf ) + um (qm ), (qf , qm )∈R2 + L reinclusions) as the instrumental variable for BA participation when the data were recorded (six months after the program s.t. qf + qm ≤ q. (9) had started). The unconditional correlation between instru- mental and instrumented variables is 84.99%. The IV results The functions uf (qf ) and um (qm ) represent the individual are reported in tables 7 to 9. The coefficients are very similar preferences of the female and male members, respectively. to the OLS estimates presented before in tables 4 to 6. The Pareto weight λ(p, w, d ) > 0 measures the female influ- We claim that the original exclusions are exogenous, so ence on the household consumption decisions. It is specified the IV estimator is consistent. We implement Hausman tests as a function of prices p ∈ RL++ , the household total income under this identification assumption. We do not reject the null w > 0, and a vector of exogenous variables d that affect hypothesis that the OLS estimator is consistent and efficient intrahousehold decision power but not individual preferences (in the class of linear estimators) for all of these regression over consumption goods. The variables in d are sometimes models. This suggests that BA participation was condition- referred to as distribution factors (DFs). Examples of DFs ally exogenous. Therefore, for the remainder of this paper, that have been used in the empirical literature are intrahouse- we focus our analysis strictly on the OLS method. hold income distribution and the wealth contributed by each member at marriage. Notice that the intrahousehold allocation derived from V. Theory Background problem (9) is Pareto optimal. Now assume for simplicity that uf and um are continuous, locally nonsatiated, and strictly We now present a theoretical background for the prediction quasi-concave. The household demand function is then given that an exogenous increase in nonlabor income accruing to a by particular household member could affect the final allocation of resources. There is a vast literature modeling household q∗ (p, w, d ) = arg max V (q, p, w, d ) s.t. p · q ≤ w. (10) expenditure decisions as resulting from different bargaining q∈RL + GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 561 Table 8.—BA Effect on Expenditure Levels: IV Method Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA dummy −3.251 0.725 −1.157 5.552∗ 3.388∗∗∗ 2.806∗ 1.334 5.959 0.808 (3.581) (5.579) (1.701) (2.848) (1.251) (1.581) (1.835) (3.738) (1.443) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 998 998 998 998 998 998 998 998 998 IV regressions with constant; dependent variables in first row; controls in first column. Instrumented variable: BA participation dummy; instrumental variable: BA initial status dummy. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Table 9.—BA Effect on Expenditure Shares: IV Method Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA dummy −0.012 −0.002 −0.003 0.004 0.007∗∗ 0.008∗∗ −0.001 0.002 −0.004 (0.008) (0.009) (0.003) (0.007) (0.003) (0.004) (0.005) (0.007) (0.004) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 998 998 998 998 998 998 998 998 998 IV regressions with constant; dependent variables in first row; controls in first column. Instrumented variable: BA participation dummy; instrumental variable: BA initial status dummy. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. An increase in female income affects the household which might have affected their consumption choices in some demand through an increase in total income (income effect). arbitrary way. But it could also influence household decisions through an In order to disentangle potential empowerment effects empowerment effect due to the increase in the female income from income and health-monitoring effects, we perform share, typically viewed as a distribution factor with impor- diff-in-diff analysis across groups of households for which tant impact on the Pareto weight. When the Pareto weight empowerment effects have different intensity. Different sub- does not depend on relative prices and DFs, the collective samples are used for that purpose in order to check the model reduces to the classical unitary model in which house- robustness of the results. hold preferences are represented by a single utility function defined on RL +. For our purpose, nonunitary models can be classified into A. Female Households two groups: one in which the Pareto weight depends on only We use 77 households with no male adult as the com- prices and total income, hence demand functions are inde- parison group in a diff-in-diff model. These households are pendent of DFs, satisfy income pooling, but may violate the typically composed of single mothers living with their chil- Slutsky symmetry condition; and another group in which the dren. No gender-specific empowerment effect should appear Pareto weight, hence the demand functions, also depends on in their consumption decisions, since females already control DFs such as the fraction of nonlabor income in the hands of all spending decisions. the female member. In the latter case, the female empower- Define Fi to be a dummy variable such that Fi = 1 indicates ment resulting from participation in the BA program should a household with no male adult (female households). The affect household consumption decisions net of pure income model in section IV is modified to be such that household i’s effects. outcome depends on Fi . When the household is enrolled in the program, VI. Identifying Female Empowerment Effects Fi Fi Fi Y1, i = μ1 + ε1,i , (11) Ideally we would like to have an experiment where the cash benefits were randomly assigned to males and females and otherwise, in different households. Unfortunately, such an experiment does not exist. Our data present some households with a Fi Fi Fi female beneficiary and others with no beneficiary. In prin- Y0, i = μ0 + ε0,i . (12) ciple, BA participation can affect household consumption patterns through two different economic channels: income We will use expenditure shares as the outcome variable Yi effects and a potential increase of the influence of the benefi- in this section. For female households, the average impact of ciary woman over household purchase decisions. In addition the BA program on Yi is given by to these two economic effects, BA beneficiaries were also required to make regular visits to public health centers, 1 − μ0 . α1 = μ1 1 (13) 562 THE REVIEW OF ECONOMICS AND STATISTICS Table 10.—Decomposing the BA Effects on Expenditure Shares Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA × Single-Female Dummy 0.010 0.056∗ −0.003 −0.029 −0.008 0.009 −0.009 −0.021 −0.004 (0.024) (0.030) (0.006) (0.025) (0.015) (0.010) (0.014) (0.026) (0.012) BA dummy −0.012∗ −0.011 −0.003 0.005 0.008∗∗∗ 0.007∗∗ 0.002 0.008 −0.004 (0.007) (0.008) (0.003) (0.006) (0.003) (0.003) (0.004) (0.006) (0.003) Single-female dummy −0.031 −0.040 −0.020∗∗ 0.043∗ 0.016 0.002 0.008 0.016 0.006 (0.025) (0.027) (0.008) (0.025) (0.016) (0.008) (0.014) (0.026) (0.012) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 1,006 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Similarly, for mixed-gender households, this average impact The program assignment is conditionally random. There- is given by fore, under the assumption that BA participation did not alter the fraction of female households, the error term υi is 1 − μ0 . α0 = μ0 0 (14) independent of the covariates in equation (18). The OLS estimates appear in table 10. Notice that the esti- The observed outcome Yi can be expressed as mates for β1 are statistically nonsignificant for all categories except family expenses. However, for this category, the sign Yi = BAi Fi Y1, 1 i + (1 − Fi )Y1,i 0 of the coefficient is positive, while it should be negative under + (1 − BAi ) Fi Y0, 1 i + (1 − Fi )Y0,i . 0 (15) the female empowerment argument. This representation leads us to B. Subsample of BE Nonparticipants Yi = β0 + β1 BAi Fi + β2 BAi + β3 Fi + υi , (16) Recall that many households that applied and were eligible for the BA program had school-aged children and were also where β0 = μ0 0 , β1 = (α − α ), β2 = (μ1 − μ0 ), β3 = 1 0 0 0 enrolled in the BE program. As in the BA program, the BE (μ0 − μ0 ). The new error term is given by 1 0 transfers are always made to the highest-ranking woman in the household. It is then possible that BE participants dif- 0,i + BAi Fi ε1,i − ε0,i − ε1,i − ε0,i υi = ε0 1 1 0 0 fered in important unobserved characteristics as they had + BAi ε0 1,i − ε0,i + Fi ε0,i − ε0,i . 0 1 0 (17) children at schooling age and the women in such households would likely have already experienced some empowerment In equation (16), the parameter β1 captures the differ- effect from the BE program. For this reason, we reestimated entiated impact of the BA transfer on female households, the previous regressions excluding BE participants from the in which the gender empowerment effect is necessarily sample. absent. Consider, for instance, the case in which income and The results are similar to the previous ones, as shown in health-monitoring effects are homogeneous across female table 11. In particular, the coefficients associated with the and mixed-gender households. Then, if female empowerment BAi Fi dummies are statistically insignificant in all regres- effects were in place, the β1 parameter should be nega- sions. Moreover, the estimated magnitude for the average tive for goods typically preferred by women. That is, the impact of the program over expenditure shares is similar to relative effect of the BA transfer over expenditure shares those previously presented. of female-specific goods should be higher in mixed-gender households in which the empowered women would have C. Subsample of Mixed-Gender Households increased influence over household decisions. As in section IV, we control for the vector Xi , which con- In many families, the eligible woman had some extra tains variables describing the number of household members, source of income apart from the BA transfer. The empower- municipality, and participation in the BE program, in order to ment effect of the BA program could potentially be different consistently identify the parameters of equation (16) through for such households compared to those with women without an OLS regression. By assuming E (υi | Xi ) = Xi · θ, the an additional source of income. We exclude female house- model becomes holds from the sample and divide the families with male and female adults into two groups: one in which the eligi- Yi = β0 + β1 Fi BAi + β2 BAi + β3 Fi + Xi · θ + υi , (18) ble woman had some extra source of income different from the BA transfer and another in which the BA transfer was the where υi = υi − E (υi | Xi ). only resource of the eligible woman. GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 563 Table 11.—Decomposing the BA Effects on Expenditure Shares: Subsample of BE Nonbeneficiaries Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA × Single-Female Dummy 0.031 0.028 −0.004 −0.018 0.012 0.005 0.003 −0.057 −0.001 (0.028) (0.043) (0.010) (0.033) (0.009) (0.014) (0.018) (0.036) (0.012) BA dummy −0.013 −0.005 −0.005 0.007 0.006∗ 0.007∗ −0.004 0.012 −0.005 (0.009) (0.011) (0.004) (0.007) (0.004) (0.004) (0.006) (0.009) (0.004) Single-female dummy −0.075∗∗ −0.001 −0.025∗ 0.057∗ −0.003 0.006 0.000 0.053 −0.012 (0.030) (0.040) (0.014) (0.033) (0.011) (0.012) (0.018) (0.037) (0.011) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 658 658 658 658 658 658 658 658 658 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Table 12.—Females with Positive Extra-Program Income Oils, Spices, and General Services Family Vices Grains Vegetables Fruits Dairy Meats Soft Drinks BA × Female-Income Dummy 0.007 −0.018 0.004 −0.009 −0.002 −0.004 0.014 0.008 0.000 (0.014) (0.016) (0.006) (0.012) (0.005) (0.006) (0.010) (0.014) (0.007) BA dummy −0.018∗ −0.006 −0.006 0.014 0.010∗∗ 0.010∗∗ −0.004 0.004 −0.004 (0.011) (0.013) (0.005) (0.009) (0.004) (0.005) (0.008) (0.011) (0.006) Female-income dummy −0.013 0.035∗∗ −0.002 −0.006 0.000 0.007 −0.012 0.001 −0.01 (0.014) (0.016) (0.006) (0.012) (0.005) (0.006) (0.010) (0.013) (0.007) Number of household members, BE, and municipality dummies Yes Yes Yes Yes Yes Yes Yes Yes Yes Household characteristics Yes Yes Yes Yes Yes Yes Yes Yes Yes Number of observations 891 891 891 891 891 891 891 891 891 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. Table 13.—Identification Test Single-Female Dummy Single-Female Dummy Female-Income Dummy (Full Sample) (BE Nonbeneficiaries) (Mixed-Gender Households) BA dummy 0.006 0.000 0.038 (0.014) (0.017) (0.026) Number of household members, BE, and municipality dummies Yes Yes Yes Household characteristics Yes Yes Yes Number of observations 1,006 658 891 OLS regressions with constant; dependent variables in first row; controls in first column. Robust standard errors in parentheses. Significant at ∗ 10%, ∗∗ 5%, ∗∗∗ 1%. This leads us to a sample with 891 households (after D. Estimation Issues excluding the 77 families with no male adult). Among them, 697 families receive the BA transfer. In 507 families, the eli- Assessing the identification assumptions. The key iden- gible woman had an extra source of income (321 of whom tification assumption behind our empirical strategy in this were BA beneficiaries). section is that participation in the BA program did not We test whether empowerment effects differ across the two affect the fraction of female households or the fraction of groups of households. (We do not have a theoretical predic- households whose women had an extra source of income. tion for sign of this effect, since that would depend on the Naturally the BA program could have affected the family sensitivity of the Pareto weight for different levels of female structure through changes in divorces and female labor sup- income.) The estimates appear in table 12, and they bring no ply.8 We show, however, in table 13 that the conditional evidence of a differentiated empowerment effect across the correlation between BA participation and the family structure two groups studied. In other words, although the BA pro- is not statistically significant. This supports the identification gram increased the consumption shares of vegetables and assumption that potential changes in family structure that fruits (and reduced the expenditure share of general services could compromise our empirical strategy have not been sta- such as utilities), this effect is not gender related. Under our tistically important during the first six months of the program maintained assumptions, this must be due to income effects (when the data were gathered). (not fully captured by the budget-share econometric represen- 8 See Rangel (2006) for evidence about changes in female labor supply tation) or to the impact of the mandatory health monitoring after a change in alimony rights and obligations to cohabiting couples in activities. Brazil. 564 THE REVIEW OF ECONOMICS AND STATISTICS Additional robustness exercises. It is important to report Angrist, Joshua D., Guido W. Imbens, and Donald B. Rubin, “Identifica- two other classes of robustness exercises that we have per- tion of Causal Effects Using Instrumental Variables,”Journal of the American Statistical Association 91 (1996), 444–455. formed. First, in order to account for potential correlation Attanasio, Orazio, and Valérie Lechene, “Tests of Income Pooling in on unobservables of different equations, we have estimated Household Decisions,” Review of Economic Dynamics 5 (2002), the entire system of budget share equations by using Zell- 720–748. Bingley, Paul, and Ian Walker, “The Labor Supply Effect of In-Kind ner’s SURE method. The estimated coefficients are identical Transfers,” Mimeograph, Lancaster University (2009). by construction, and the confidence intervals do not change Bradbury, Bruce, “Consumption and the Within-Household Income Distri- considerably. The qualitative results are identical to those bution: Outcomes from an Australian ‘Natural Experiment,”’ CESifo Economic Studies 50 (2004), 501–540. presented before. Browning, Martin, “Children and Household Economic Behavior,”Journal Moreover, we have also included the household total of Economic Literature 30 (1992), 1434–1475. expenditure as a control variable in OLS regressions equiv- Browning, Martin, François Bourguignon, Pierre-André Chiappori, and Valérie Lechene, “Income and Outcomes: A Structural Model alent to those presented in tables 5 to 6 and 8 to 12. Under of Intrahousehold Allocation,” Journal of Political Economy 102 the strong assumption that total expenditure is independent (1994), 1067–1096. of an observable characteristics of the household, this exer- Browning, Martin, Pierre-André Chiappori, and Valérie Lechene, “Collec- tive and Unitary Models: A Clarification,” Review of Economics of cise would disentangle potential income effects (captured by the Household 4 (2006), 5–14. the total expenditure variable) from health-monitoring effects Chen, Zhiqi, and Frances Woolley, “A Cournot-Nash Model of Family (which would remain captured by the BA dummy). The quali- Decision Making,” Economic Journal 111 (2001), 722–748. Chiappori, Pierre-André, “Rational Household Labor Supply,” Economet- tative results are identical—that is, the main impact of the BA rica 56:1 (1988), 63–90. program (conditional on total expenditure) was an increase Costa, Dora L., “Displacing the Family: Union Army Pensions and Elderly in the budget share allocated to vegetables and fruits and Living Arrangements,” Journal of Political Economy 105 (1997), 1269–1292. a decrease in the share allocated to general services (such Duflo, Esther, “Child Health and Household Resources: Evidence from as utilities). This suggests that compliance with vaccination the South African Old Age Pension Program,” American Economic schedules and regular visits to public centers for prenatal care, Review: Papers and Proceedings 90 (2000), 393–398. ——— “Grandmothers and Granddaughters: Old Pension and Intra- child-growth monitoring, and health and nutrition classes Household Allocation in South Africa,” World Bank Economic may be behind the observed increase in the consumption of Review 17:1 (2003), 1–25. healthy foods, such as vegetables and fruits. Edmonds, Eric V., Kristin Mammen, and Douglas L. Miller, “Rearranging the Family?” Journal of Human Resources 40:1 (2005), 186–207. Fiszbein, Ariel, Norbert Schady, Francisco Ferreira, Margaret Grosh, Niall Kelleher, Pedro Olinto, and Emmanuel Skoufias, Conditional Cash VII. Conclusion Transfers: Reducing Present and Future Poverty (Washington, DC: World Bank, 2009). We study gender bias in the intrahousehold allocation of Handa, Sudhanshu, “Expenditure Behavior and Children’s Welfare: An resources using data from a Brazilian social program, Bolsa Analysis of Female Headed Households in Jamaica,” Journal of Development Economics 50 (1996), 165–187. Alimentação (BA). The BA program was designed to reduce Heckman, James J., Hidehiko Ichimura, and Petra Todd, “Matching as an nutritional deficiencies and infant mortality among the poor- Econometric Evaluation Estimator,” Review of Economic Studies 65 est households in Brazil. It relies on demand-side incentives (1998), 261–294. Hotchkiss, Julie L., “Do Husbands and Wives Pool Their Resources? by means of money transfers to pregnant women and mothers Further Evidence,” Journal of Human Resources 40 (2005), 519– of young children in low-income families. Due to bureau- 531. cratic mistakes, many eligible applicants did not receive the IFPRI, Estudo de Avaliação de Impacto para a Bolsa Alimentação (Wash- ington, DC: International Food Policy Research Institute, Food cash benefit. These unintentional exclusions formed a control Consumption and Nutrition Division, 2003). group in the mold of a random experiment. Imbens, Guido W., and Jeffrey M. Wooldridge, “Recent Developments We do not find evidence that household consumption deci- in the Econometrics of Program Evaluation,” Journal of Economic Literature 47:1 (2009), 5–86. sions were affected by the fact that the program transfer was Lindert, Kathy, Emmanuel Skoufias, and Joseph Shapiro, “Redistributing directed to a woman instead of a man. It is important to stress Income to the Poor and the Rich: Public Transfers in Latin Amer- that the households in our sample are very poor and spend ica and the Caribbean,” World Bank SP discussion paper no. 0605 (2006). most of their income on basic goods. Therefore, our results Lundberg, Shelly J., and Robert A. Pollak, “Separate Spheres Bargaining are not representative of how middle-class households react and the Marriage Market,” Journal of Political Economy 101 (1993), to exogenous changes in female nonlabor income. Instead, 988–1010. Lundberg, Shelly J., Robert A. Pollak, and Terence J. Wales, “Do Husbands they contribute to the literature on development economics and Wives Pool Resources? Evidence from the UK Child Benefit,” by questioning the growing belief that CCT gender-oriented Journal of Human Resources 32 (1997), 463–480. transfers affect intrahousehold expenditure decisions. Manser, Marilyn, and Murray Brown, “Marriage and Household Decision Theory: A Bargaining Analysis,” International Economic Review 21:1 (1980), 31–44. REFERENCES McElroy, Marjorie B., and Mary J. Horney, “Nash-Bargained Household Decisions: Towards a Generalization of the Theory of Demand,” Abadie, Alberto, and Guido W. Imbens, “Large Sample Properties of Match- International Economic Review 22 (1981), 333–349. ing Estimators for Average Treatment Effects,” Econometrica 74:1 Morris, Saul S., Pedro Olinto, Rafael Flores, Eduardo A. F. Nilson, and (2006), 235–267. Ana C. Figueiró, “Conditional Cash Transfers Are Associated with ——— “Estimation of the Conditional Variance in Paired Experiments,” a Small Reduction in the Rate of Weight Gain of Preschool Children Annales d’Economie et de Statistique 91 (2008), 175–187. in Northeast Brazil,” Journal of Nutrition 114 (2004), 2336–2341. GENDER BIAS IN INTRAHOUSEHOLD ALLOCATION 565 Pezzin, Liliana E., and Barbara S. Schone, “The Allocation of Resources ——— “Matched Sampling for Causal Effect” (Cambridge: Cambridge in Intragenerational Households: Adult Children and Their Elderly University Press, 2006). Parents,” American Economic Review Papers and Proceedings 87 Thomas, Duncan, “Intra-Household Resource Allocation: An Inferential (1997), 460–464. Approach,” Journal of Human Resources 25 (1990), 635–664. Rangel, Marcos A., “Alimony Rights and Intrahousehold Allocation of ——— “Like Father, Like Son; Like Mother, Like Daughter: Parental Resources: Evidence from Brazil,” Economic Journal 116 (2006), Resources and Child Height,” Journal of Human Resources 29 627–658. (1994), 950–988. Rosenbaum, Paul. R., and Donald B. Rubin, “Constructing a Control Group Thomas, Duncan, Dante Contreras, and Elizabeth Frankenberg, “Distribu- Using Multivariate Matched Sampling Methods That Incorporate the tion of Power within the Household and Child Health,” mimeograph, Propensity Score,” 39:1 (1985), 33–38. UCLA (2002). Rubin, Donald B., “Using Multivariate Matched Sampling and Regression Ward-Batts, Jennifer, “Out of the Wallet and into the Purse: Using Micro Adjustment to Control Bias in Observational Studies,” Journal of Data to Test Income Pooling,” Journal of Human Resources 43 the American Statistical Association 74 (1979), 318–328. (2008), 325–351.