investments, tp and tranfers made to women

Investments, Time Preferences, and Public Transfers Paid to WomenAuthor(s): LuisRubalcava /GracielaTeruel /DuncanThomasSource: Economic Development and Cultural Change, Vol. 57, No. 3 (April 2009), pp. 507-538Published by: The University of Chicago PressStable URL: http://www.jstor.org/stable/10.1086/596617 .Accessed: 12/05/2015 17:13

Your use of the JSTOR archive indicates your acceptance of the Terms & Conditions of Use, available at .http://www.jstor.org/page/info/about/policies/terms.jsp

.

JSTOR is a not-for-profit service that helps scholars, researchers, and students discover, use, and build upon a wide range ofcontent in a trusted digital archive. We use information technology and tools to increase productivity and facilitate new formsof scholarship. For more information about JSTOR, please contact [email protected].

.

The University of Chicago Press is collaborating with JSTOR to digitize, preserve and extend access toEconomic Development and Cultural Change.

http://www.jstor.org

This content downloaded from 200.10.244.14 on Tue, 12 May 2015 17:13:35 PMAll use subject to JSTOR Terms and Conditions

2009 by The University of Chicago. All rights reserved. 0013-0079/2009/5703-0001$10.00

Investments, Time Preferences, and Public TransfersPaid to Women

luis rubalcavaCentro de Analisis y Medicion del Bienestar Social A.C. and Centro de

Investigacion y Docencia Economicas A.C.

graciela teruelUniversidad Iberoamericana

duncan thomasDuke University

I. Introduction

Research suggests that men and women do not share the same preferences.First, in carefully controlled experimental settings, women have been shownin many instances to be more altruistic and more risk averse than men (Eckeland Grossman 1998; Andreoni and Vesterlund 2001; see Eckel and Grossman[2008a, 2008b] for reviews). However, the populations in most of these studiesare college students, and the generality of the evidence has not been established.Second, nonexperimental evidence, based on population surveys, suggests thatin some contexts women allocate resources under their control toward goodsthey or their children consume (such as clothing; see Lundberg, Pollak, andWales 1997) and also to investments that improve child health and well-being(e.g., Thomas 1990; Duflo 2000). This paper addresses two shortcomings inthis literature. First, legitimate concerns have been raised regarding the extentto which this evidence is contaminated by unobserved heterogeneity that is

We thank Robert Pollak, John Strauss, Alessandro Tarozzi, and three referees for valuable com-ments and Rodolfo Islas and Armando Baigts for outstanding research assistance. Financialassistance from the Eugene Kennedy Shriver National Institute of Child Health and HumanDevelopment (R01HD047522, R21HD049766, and R01HD050452), the Fogarty InternationalCenter (D43TW007699), and the Hewlett Foundation/Population Reference Bureau is gratefullyacknowledged.


508 economic development and cultural change

correlated with the distribution of resources within households.1 Second, directmeasures of preferences of men and women provide an opportunity to evaluateinterpretations of the evidence on household resource allocations.

To address the first concern, we exploit variation in the distribution ofresources within households induced by a social experiment in which assign-ment to the treatment group is random and women in the treatment house-holds are given income. PROGRESA, one of the most ambitious antipovertyprograms in the world, provides cash transfers to poor rural households inMexico, and these transfers are paid to women. The payment is large: onaverage, beneficiary households receive payments that are around one-quarterof pretreatment household income. We compare the impact on allocationdecisions of PROGRESA income with the impact of income from all othersources in order to separately identify the effect of a greater share of householdincome in the hands of women from other effects of the PROGRESA programon resource allocation decisions. In addition to comparing treatment withcontrol households, we focus on PROGRESA beneficiaries and exploit variationin both income and the timing of PROGRESA benefit payments.

Relative to other income, PROGRESA income tends to be spent on in-vestments in the future. Specifically, resources are allocated toward child goodsand toward investment in small livestock, which, in the study societies, aretraditionally cared for by (and under the control of ) women. The results arerobust to focusing on variation in the timing of PROGRESA payments withintreatment households and also to controlling expected future benefits. Thissuggests that it is actual income in the hands of women that affects within-

1 Thomas (1990) compares nonlabor income of males and females; women with relatively morenonlabor income may be different from other women in other dimensions including, e.g., timepreferences, which are related to saving and investment decisions. Lundberg et al. (1997) exploita natural experiment in the United Kingdom in which Child Benefit was paid to women ratherthan to men. Hotchkiss (2004) argues that the evidence can be explained by a time effect andnotes similar budget reallocations among couples who did not receive the benefit, presumablybecause of changes in relative prices coincident with the change in the way the benefit was paid.Noting that, in the United States, Aid to Families with Dependent Children (AFDC) was paidonly to single women with children, Rubalcava and Thomas (2000) point out that AFDC provideda potential safety net for a lower-income married woman with children in the event of separationfrom her partner. Variation in benefit levels across states and over time provides natural variationin the womans bargaining power even while she is married and is shown to affect householdresource allocations. Duflo (2000) exploits a different natural experiment in which older adultswere given pensions in South Africa. She finds that children are healthier in households with olderwomen. Edmonds, Mammen, and Miller (2005) demonstrate that household composition respondsto receipt of the pension and that young children are more likely to coreside with older womenwho are eligible for the pension. Behrman, Pollak, and Taubman (1986) and Behrman and Ro-senzweig (2004) provide evidence on within-family resource allocations under different assumptions.


Rubalcava, Teruel, and Thomas 509

household resource allocation decisions. If the evidence reflects the impact ofunobserved heterogeneity in the communities that received the treatment orthe effect of other dimensions of the program, PROGRESA income shouldhave the same impact on single-headed households that received the benefits.It does not. Among those households, PROGRESA income has the sameimpact on spending and savings decisions as other sources of income. Weconclude that our results for couple households are unlikely to be driven byunobserved heterogeneity.

An implication of this evidence is that if receipt of PROGRESA incomeincreased the bargaining position of women within poor, rural Mexican house-holds, then those women have longer time horizons than men when planningresource allocations. We investigate the latter hypothesis drawing on directmeasures of intertemporal preferences collected from a different sample of ruralhouseholds in the Mexican Family Life Survey (MxFLS). These measures in-dicate that women are more patient than men. This is the first instance inwhich evidence on the relationship between control over resources and theirallocation within households has been directly linked to specific domains ofpreferences.

We conclude that among poor rural Mexicans, women have longer planninghorizons, and so resources under their control, including PROGRESA income,are likely to be spent on investments in children and in small-scale livestock.The latter, at least, typically remain under the control of women over time,which will likely contribute to further enhancing their relative bargainingposition within the household.

Section II describes the design of the PROGRESA program. The modelmotivating our research precedes a description of the data. Issues that areconfronted in the empirical implementation are discussed and followed byresults of data from the PROGRESA study. Section VII presents evidence onintertemporal preferences from the MxFLS.

II. The PROGRESA ProgramPROGRESA, the centerpiece of the Mexican governments antipoverty strat-egy, is a conditional income transfer program that began in 1997 in ruralareas and was subsequently expanded into urban areas when it was renamedOportunidades.2 The program covers nearly 5 million families, which is almosta quarter of the Mexican population. Arguably the most ambitious conditional

2 We refer to PROGRESA rather than Oportunidades since we use data on rural households priorto the expansion of the program to urban areas. See Parker, Rubalcava, and Teruel (2008) forfurther description of the program.



income transfer program in the world, PROGRESA serves as a model forsimilar programs throughout Latin America and the Caribbean.

The average eligible household is given an income transfer of around 30pesos per person per month.3 This is a very large transfer, which amounts toover 28% of average monthly per capita expenditure of these households. Thevalue of the transfer depends on whether household members aged 22 andyounger attend school4 and whether all household members attend the localpublic health clinic.5

Key for this study is that, whenever possible, all benefits are paid directlyto women, typically the mothers of age-eligible children, who pick up thepayment at the local post office. (If there is no eligible woman in the household,the payment is made to a male adult.) This design was motivated by a beliefamong the program architects that giving income to women would be moreeffective in increasing investment in the next generation and reducing povertythan giving income to men. This paper subjects that belief to empirical scrutinyby examining the impact of the income transfer on household resource allo-cations.

PROGRESA is means tested with a two-stage targeting mechanism. First,communities that are deemed poor (on the basis of socioeconomic character-istics) are selected. Second, the ENCASEH, a census of all households in thecommunity, is conducted; a household is eligible for PROGRESA if it fallsbelow a multidimensional poverty cutoff (as measured by a combination ofincome, demographic characteristics, educational attainment of householdmembers, the presence of disabled individuals in the household, housing char-acteristics, and the ownership of durable goods, animals, and land).6

The list of eligible households is announced at a meeting in the communityto build consensus that the selection mechanism is fair. In practice, this last

3 One peso was worth US$0.11 in 1997.4 The grant is increased by 70 pesos for each child who attends the third grade of primary school.The amount is increased with grade completion. For example, it is increased by 225 pesos and255 pesos for boys and girls in the third grade of secondary school, respectively. If a child missesmore than three school days in a month (for unjustified reasons), the household does not receivethe grant that month.5 Basic, preventive health care services are provided by the public sector for all household members.Benefits are paid only if household members attend health clinics on a schedule spelled out by theprogram. In addition, households are given 145 pesos per month for food in addition to nutritionsupplements, which are principally targeted to children between the ages of 4 months and 2 yearsand pregnant and lactating women. The school attendance of children and family health visits areverified through school and clinic records.6 See Skoufias, Davis, and Behrman (1999) for a description and evaluation of the targetingmechanism.



step rarely results in substantial changes to the list of eligible families. Eli-gibility is fixed after the initial assignment.

III. Theoretical FoundationIn order to motivate the empirical strategy, we lay out a simple model ofhousehold behavior that provides a set of testable hypotheses regarding theeffect of PROGRESA income on household resource allocations. We thenproceed to discuss our empirical strategy and the assumptions that are neededin order to interpret the results.

Begin with a model of household behavior in which the well-being of allhousehold members in any period t, , depends on the utility of each member,Wt

. In turn, each individuals utility, , depends on the com-mp 1, ... , M Umtmodity consumption of all household members, , , where gx gp 1, ... , Ggmtindexes goods, and let denote consumption of leisure of each individualx0mtat time t. We allow tastes, and therefore utility, to be affected by individual-and household-specific characteristics. Let denote those that are observed,mtsuch as household demographic structure and socioeconomic status, and let

represent all unobserved characteristics, such as tastes for work, consumption,taltruism, risk, and intertemporal preferences. Each individuals subutility func-tion is given by , which is assumed to be quasi-concave, non-U (x , m , )mt t t tdecreasing, and strictly increasing in at least one argument. The householdwelfare function, W, aggregates these individual subutility functions:

WpW[U (x , m , ), ... , U (x , m , ), l (p, y)], (1)t t 1t t t t Mt t t t mt

where the weights can be interpreted as distribution factors (Browninglmtet al. 1994; also called extra environmental parameters by McElroy [1990]),which have no direct effect on an individuals utility but affect decision makingwithin the household. Intuitively, indicates member ms bargaining powerlmtwithin the household at time t and depends on past, present, and futureindividual, household, and environmental characteristics. These include ob-served characteristics, p, such as age, education, income, and wealth of eachhousehold member as well as prices, interest rates, marriage market oppor-tunities, customs, laws, and institutions that affect marriage and divorce atthe community or society level. The distribution factors may also depend onunobserved characteristics, y, such as attitudes toward risk, altruism, trust,and intertemporal preferences. Note that p and y, which may vary acrossindividuals and over time for a particular individual, will, in general, includefactors that affect each individuals utility as well as resources available to thatperson and to the household. The model is quite general and includes notonly resource allocations that are Pareto efficient (Chiappori 1988) but also



noncooperative outcomes including models in which household members haveseparate spheres of interest (Lundberg and Pollak 1993; Duflo and Udry 2004).

The household welfare function (1) is maximized subject to the intertem-poral household budget constraint:

A A p (1 r )A (1 r )A mt 0t mt mt1 0t 0t1m m

t p (T x ) p x . (2) mt 0mt 0mt t t[ ]m

In period t, household assets are given by the sum of the assets of eachmember, , and jointly owned assets, . They are equal to assets in theA Amt 0tprior period, the return on those assets, r, plus savings, which is given byincome less expenditure. Income comprises transfer income plus earnings.Transfer income, t, is net income from private transfers (with noncoresidentfamily members, e.g.) plus public transfers (which includes PROGRESA).Earnings of member m are the product of the wage, , and the amount ofp0mttime spent working, which is the total amount of time, T, less the amountof time spent not working, . All prices other than wages are assumedx p0mt tto be taken as given by household members. The return on assets, r, is allowedto be individual idiosyncratic, which would arise, for example, if marketopportunities differ for men and women because of restrictions on behavior.

Unitary Model of the HouseholdThe simplest model of the household, which is widely used in the socialscience literature, assumes that all household members behave as if the house-hold were a single decision-making unit. This will arise if the distributionfactors, , are fixed over time. It may also arise if one household member,lmta dictator, makes all allocation decisions, in which case the distribution factors,

, are zero for all but that member and the aggregator function reducesl W(7)mtto that members subutility function. An observationally equivalent assump-tion is that all subutility functions in (1) are identical. Under any of theseassumptions, the household may be treated as if it were a single unit so thatthere is no place for dissension within the household and, therefore, for anyindividual to assert his or her power in decision making. While this modelis clearly a simplification, it has proved to be extremely powerful as an or-ganizing principle in the theoretical and empirical literature on householdand family decision making.

In this model, decisions about spending on goods and services, savings,7

7 To keep notation simple, savings is treated as spending on investment goods or assets.



and time allocation in any period depend on total household income, ymt(which includes the return on assets, transfers, and earnings); household char-acteristics, m, such as permanent wealth and sociodemographic composition;prices, p; and factors such as tastes, :

M

x p x y , m, p , . (3)( )gt gt mt t gt0

In a life cycle model with no liquidity constraints and no uncertainty, currentspending will not depend on current income. That model has been widelyrejected in the literature, and so the restriction is not imposed here. For ourpurposes, the key point in this model is that saving and spending patternsare not influenced by who within the household receives the income or ownsthe assets. If (3) is a good approximation of demand functions for poor house-holds in rural Mexico, then it will not matter a whit whether PROGRESAincome is paid to the mother, to the father, or to anyone else. That hypothesiswill be tested below. Prior to laying out our testing strategy, it is useful tospell out a class of models in which resources of individuals do affect householdchoices in order to demonstrate that this test has power against plausiblealternatives.

Individualistic Models of the HouseholdThe most primitive model of behavior treats the individual as the primaryelement in decision making, with the household simply serving as a structure,such as a club or group, in which decisions are aggregated. There is a wideclass of individualistic models in the literature. A simple, quite general modelassumes that allocation decisions are the outcome of some repeated game thatcan be approximated as achieving a cooperative equilibrium and that allocationsare Pareto efficient. This is an intuitively appealing assumption when thinkingabout the behavior of household members who share much in common andare likely to be altruistic toward one another (Chiappori 1988, 1992, 1997;Bourguignon and Chiappori 1992; Browning and Chiappori 1998, 2000; seeBourguignon, Browning, and Chiappori [1995, 2006] for a fuller discussion).Other models involve noncooperative equilibria and highlight the role ofspheres of interest in bargaining or separate purses of husbands and wives (seeManser and Brown 1980; McElroy and Horney 1981; Lundberg and Pollak1993; Duflo and Udry 2004).

Presumably the reason that individuals form a household is that it producesgoods and services for its members that they would not be able to consumeif they were not organized in the household. These may be the benefits as-



sociated with altruism and caring, with returns to scale in the production ofgoods and services such as meals or housing or externalities that householdsprovide. The benefits may also be manifest in resolving coordination andinformation asymmetries. In both the cooperative and noncooperative for-mulations of the model, what the household produces and who benefits fromthat production depend on the power a member wields in asserting his or herpreferences over those of others. There are many ways in which power maybe manifest, and it may depend on such factors as the options one might haveif one left the household. If the power of each household member is denotedby the vector of weights , then spending and savings decisions on any goodlgtg in any period are given by

M

x p x l , y , m, p , . (4)( )gt gt gt mt t gt0

Apart from the weighting factors, l, the demand functions in this individ-ualistic model, (4), are identical to those under the assumptions of the unitarymodel, (3).

The weights play a central role in the model and reflect the relative im-portance of each members power in affecting household allocation decisions.If allocations are Pareto efficient, then the weights will not vary across goods.This places restrictions on how power affects resource allocations. In general,the weights will likely respond to changes in the relative power of householdmembers induced, for example, by programs that are targeted toward onegroup of people rather than another. PROGRESA is designed to be such aprogram.

In general, estimation of (4) is complicated for at least two reasons. First,in studies based on observational data, it is not clear how to measure changesin power. We exploit the fact that communities are randomly assigned to atreatment or control group in the PROGRESA evaluation, and only householdsin treatment communities receive PROGRESA income. Since this income ispaid to women, resources in the hands of women in the treatment group willhave increased whereas resources will not have changed in the control group.Since total household income will also be higher in treatment households, itis possible that we will assign a power effect to what is, in fact, an incomeeffect. As explained in detail below, we address this issue by relying on acomparison of the marginal effect on spending patterns of PROGRESA incomewith the marginal effect of other household income.

A second complex issue in this literature revolves around the fact that themajority of studies proxy power with the distribution of earnings within the



household. That distribution reflects current (and previous) decisions aboutwork and savings, and those decisions are likely to be related to unobservedcharacteristics of household members that also affect resource allocations. Forexample, if a woman wishes to invest more in her children, she may seek outearnings opportunities and spend disproportionately more of those resourceson her children. She may also invest more of her time and energy in herchildren, and those investments would, in general, be captured by in (4).gtIn that case, the distribution of earnings and unobserved characteristics in theregression will be correlated and estimates of the effects of individual earningswill be biased. A key advantage of examining the behavior of households thatreceive PROGRESA benefits is that income is paid to households on the basisof their socioeconomic and demographic characteristics at the time of enroll-ment into the program. The benefit does not respond to changes in (non-PROGRESA) household income or labor supply of household members thatmight occur after program enrollment. Conditional on all observed and un-observed characteristics at the time of enrollment, the receipt of PROGRESAincome can, therefore, be treated as an exogenous shift in the distribution ofcontrol over resources within the household.8

The next section describes the PROGRESA data. We then present ourempirical strategy and explain how the experimental design of PROGRESAis exploited to test the predictions of the unitary model.

IV. Data

An important dimension of the design of PROGRESA for the purposes ofthis study is the fact that the government was committed to conducting arigorous evaluation of the impact of the program. In 1997, prior to theintroduction of the program, 506 communities in seven states9 were selectedfor the rural evaluation sample. In May 1998, 63% of the communities wereassigned to receive PROGRESA benefits (treatment communities) and the restwere designated to be phased in to PROGRESA 3 years later, toward the endof 2000 (control communities). In 1998, program officials announced to alltreatment households that the benefits would be paid for at least 3 years.Control households were not notified about the program. (In fact, as households

8 It is possible that households respond to the program by changing labor supply (or time allocation),transfers in or out of the household, or shifting type of work, crop choice, or technology choice.Under the null that the unitary model is correct, these choices are made at the household leveland will not reflect the preferences of individuals within the household. To the extent that suchbehavioral responses do not change after the initiation of benefit payments, they are addressed inthe empirical analyses below.9 These states were among the first states to receive PROGRESA benefits.



TABLE 1DISTRIBUTION OF HOUSEHOLDS IN 1997 BASELINE

Number of Households Treatments Controls Total

Not eligible (not poor) 7,003 (61%) 4,531 (39%) 11,534 (48%)Eligible (poor) 7,830 (63%) 4,678 (37%) 12,508 (52%)

Total 14,833 (62%) 9,209 (38%) 24,042

Source. 1997 baseline survey (ENCASEH).Note. Row percentages are in parentheses.

TABLE 2CHARACTERISTICS OF ALL HOUSEHOLDS HEADED BY A COUPLE

Eligible forPROGRESA

(1)

Not Eligiblefor PROGRESA

(2)

Heads years of schooling 2.87 3.06(.06) (.06)

Age of head 41.83 50.80(.20) (.29)

Household size 6.00 4.77(.04) (.05)

Number of households10,694 8,806

Eligible Households Headed by a Couple

Treatments(1)

Controls(2)

Difference(3)

Heads years of schooling 2.91 2.81 .10(.08) (.10) (.13)

Age of head 41.73 42.00 .27(.24) (.35) (.42)

Household size 5.99 6.05 .06(.05) (.06) (.07)

Number of households 6,683 4,011 10,694

Source. 1997 baseline survey (ENCASEH).Note. Robust standard errors are in parentheses.

in control communities became aware of the program, pressure to includethem in the program mounted. The communities started receiving benefitsin early 2000.)

On the basis of data from a census of over 24,000 households conductedin late 1997 in all the PROGRESA evaluation sample communities, com-munities were matched in terms of propensity scores based on levels of in-frastructure and economic status. Two communities in each triple were ran-domly assigned to the treatment group, and the third was assigned to thecontrol group (Behrman and Todd 1999). Table 1 reports the distribution ofhouseholds in the baseline census. Slightly over 50% were eligible for PRO-GRESA (which we shall call poor for shorthand), and about two-thirds ofthe households were in treatment communities.

Table 2 summarizes some demographic characteristics of all households



headed by a couple.10 The upper section compares those eligible for the PRO-GRESA benefit (col. 1) with those not eligible (col. 2). Households eligiblefor the PROGRESA benefit are, by design, poor. Relative to other households,they are earlier in the life course and have more members, and the head hasless education. The lower section of the table compares treatment with controlhouseholds among those eligible for the benefit. Since communities wererandomly assigned to the treatment, there should be no differences in so-ciodemographic characteristics of the two groups. The differences are small,and none is significant.

After the baseline, follow-up surveys of all treatment and control householdswere conducted about every 6 months until 2000. Detailed expenditure, in-come, and asset data were collected from each household in the follow-upsurveys in March 1998, October 1998, May 1999, and November 1999(ENCEL).11 Table 3 summarizes data drawn from these three surveys for treat-ment households (col. 1), control households (col. 2), and the difference (col.3). (Unfortunately, no expenditure, income, or asset data were collected in thebaseline census.)12

Household per capita expenditure is reported in the first row of the table.On average, treatment households spend about 15 pesos per person per monthmore than control households. In part, this reflects that treatment householdsreceived the PROGRESA benefit. The household survey data have beenmatched with the value and timing of each payment to each beneficiaryhousehold reported in PROGRESAs administrative records. In this sample,the average PROGRESA benefit is slightly over 30 pesos per capita per month,indicating that treatment households must be saving part of the benefit. Thisis reflected in the second row of panel A of the table, which indicates thattreatment households report saving over 13 pesos per capita per month morethan controls.

10 The majority of the analyses reported below are restricted to households headed by a couple inevery wave of the surveys. They account for 95% of all households. Moreover, dissolution rates arethe same for treatment and control households. Results for single-headed households provide usefulchecks on the assumptions and are also reported below.11 Note that in effect there are two baseline surveys that can be used as part of the evaluation: theENCASEH and the ENCEL March 1998 survey. Neither survey collected information on householdexpenditures and household animal ownership. This study relies, therefore, on postprogram house-hold resource allocation data.12 Since no attempt was made to follow movers, attrition is potentially a concern for the inter-pretation of the results. While one-third of households left the sample during the study period,the key for our purposes is whether attrition differs depending on treatment/control status. Forcouple households, it does not, and this is true even after we control household characteristics andthe PROGRESA eligibility criteria. See Teruel and Rubalcava (2006) for a general discussion ofattrition in these data.



TABLE 3SAVINGS, OWNERSHIP OF LIVESTOCK, NUTRIENT INTAKE, AND BUDGET SHARES

Treatments(1)

Controls(2)

Difference(3)

A. Expenditure, savings, and livestock:Household expenditure per capita

(monthly) 133.39 118.23 15.16(.77) (.93) (1.21)

Household income: expenditure per capita(monthly) 157.88 144.49 13.38

(14.89) (19.31) (24.38)Number of chickens and turkeys 4.45 4.10 .36

(.05) (.06) (.08)Number of pigs .79 .80 .02

(.01) (.02) (.02)Number of cows .44 .34 .10

(.02) (.02) (.03)Number of horses and donkeys .46 .39 .07

(.01) (.01) (.01)B. Nutrient intake:

Calories per capita 1,807 1,714 94(11.54) (14.10) (18.22)

Protein per calorie (g/Kcal) 2.39 2.32 .06(.05) (.10) (.10)

C. Budget shares on food:Food 66.69 66.68 .01

(.19) (.24) (.31)Vegetables 9.57 8.97 .59

(.06) (.07) (.09)Fruits .56 .41 .16

(.01) (.02) (.02)Tortillas and beans 15.31 17.79 2.48

(.11) (.17) (.20)Meat 12.19 10.56 1.64

(.09) (.11) (.14)D. Other budget shares:

Education 1.58 1.55 .02(.04) (.05) (.06)

Boys clothing 2.15 1.62 .54(.03) (.03) (.04)

Girls clothing 1.97 1.46 .50(.03) (.03) (.04)

Observations 14,413 8,469

Source. ENCEL, October 1998, May 1999, and November 1999.Note. Robust standard errors are in parentheses.

Few rural Mexican households have any financial savings, but many ownsome livestock, which provide a key mechanism through which householdsmay save.13 The surveys record the number of livestock owned by the household

13 See, e.g., Arriaga-Jordan and Pearson (2004), who note that livestock is a major source ofsavings (105) as well as a source of future income through output (eggs, meat, and milk), by-products (manure and foraging), and services (draught power).



in each of several categories. The remainder of panel A demonstrates thattreatment households own significantly more cows and more horses and don-keys and particularly more chickens and turkeys. We estimate that thesedifferences account for about 70% of the difference in reported saving oftreatment households relative to controls, cumulated over the 18 months sincethe inception of PROGRESA.14

Livestock are particularly interesting in the context of our research questiongiven the ethnographic literature, which has shown that in rural Mexico, asin many low-income societies, women are more involved in small-scale sub-sistence livestock-rearing [such as poultry and pigs] and men are more likelyto be involved in large scale, cash-generating production such as cattle, horses,and donkeys (von Keyserlingk 1999, 41).15 Moreover, poultry are often con-sumed by household members, and so these investments are likely to contributeto improving the nutritional status of household members in the future.

As panel B indicates, treatment households consume more nutrients andhigher-quality nutrients. Specifically, relative to controls, on average, individ-uals in treated households consume almost 100 calories more per day, and thecalories they consume are of higher quality (as measured by protein per cal-orie).16 As displayed in panel C, the higher-quality diets of treatment house-holds are also reflected in higher per capita spending on food (since food sharesare the same and treatment households have higher per capita spending overall).Moreover, relative to controls, treatment households allocate more of the bud-get toward meat and vegetables and to a less extent fruit, with less beingallocated to tortillas and beans.

Finally, panel D of the table reports the average share of the budget spenton investments on children. The evidence shows that treatment householdsallocate more resources to their children, possibly as another way to save forthe future. The PROGRESA benefit is higher if age-eligible household mem-bers attend school; and if schooling incurs costs, then higher spending maysimply reflect the additional costs of schooling. Treatment households allocateessentially the same fraction of the budget to schooling as controls. Thistranslates into higher spending on schooling by treatment households. The

14 In the November 1998 wave of the evaluation survey, households reported the quantity andvalue of livestock categories. For this calculation, livestock have been evaluated using the unitvalues for each category averaged over the entire sample. The number of livestock is used in theanalyses below because no information on values is recorded in the other rounds of the survey.15 Arizpe, Botey, and Deere (1986, 74) comment that some duties are considered exclusivelyfeminine . . . taking care of poultry and, sometimes, pigs, whereas men are responsible for feedingand grazing cattle and horses.16 Nutrient intakes are computed by converting quantities of food consumed into calories andprotein using standardized food tables for Mexico (Perez and Marvan 2001).



difference, however, is small (about 0.25 peso per month per capita), and itis only marginally significant. The lions share of additional expenditure amongtreatment households on children is spent on their clothing. Treatment house-holds allocate significantly more of their budget to childrens clothing, andthese differences amount to about 2.5 pesos per month.17

In sum, the average treatment effects in table 3 suggest that treatmenthouseholds allocate PROGRESA income to investments that are likely toimprove the well-being of household members in the future. These includeinvesting in livestock, more spending on children, and higher-quality diets.Since the benefit was paid to women, it is tempting to interpret these effectsas indicative of the impact of empowering women. That interpretation wouldbe premature. As is clear in (4) above, the PROGRESA benefit has an incomeeffect (because total resources available to the household are increased) andmay also affect the distribution of power, l, within the household. Both ofthese are reflected in the estimated average treatment effect. We turn next toa regression framework in an effort to separate these effects.

V. Regression Results

A naive interpretation of (4) suggests that when total household resources arecontrolled, differences in allocations by treatment and control households maybe interpreted as a rejection of the unitary model since the differences indicatethat PROGRESA income affects demand by shifting the distribution of power,l, within the household over and above the income effect. However, PRO-GRESA provides beneficiaries with a package of support that includes notonly income but also incentives for children to attend school, incentives forall household members to attend health clinics, and a modest food supplement.These additional components of the PROGRESA intervention likely influencethe production of human capital and may therefore directly affect allocationdecisions within treated households. For example, nutrition counseling is pro-vided at health clinics, which may result in households shifting resources toimproved nutrition. By using only information on participation in the pro-gram, it is not possible to separate the impact of the income transfer onspending and saving decisions from the direct effects due to these additionalcomponents of PROGRESA.

We therefore follow a different approach and examine the marginal effectof PROGRESA income on allocations, controlling total household resources

17 The budget shares in the table are not exhaustive. Conditional on total household resources,PROGRESA income has no effect on the shares on several commodity subgroups, including health,personal care, household goods, and semidurables.



(including PROGRESA income). If the marginal effect is zero, then PRO-GRESA income has the same impact as any other income and the unitarymodel is not rejected. If the marginal effect is not zero, we interpret it as theimpact of an exogenous increase in the share of household resources that areunder the control of women, relative to men, which operates through l in(4). Relative to a comparison of the average spending of treatment and controlhouseholds, this is a substantially more subtle test of the effect of changingthe distribution of resources within the household than the average treatmenteffects in table 3. The interpretation is more complicated if the receipt ofPROGRESA income affects other sources of income. We have explored theissue and find no evidence of differences in labor earnings or net private transfersbetween treatment and control households in our sample during the studyperiod.18

Results from estimates of model (4) are reported in table 4, which presentsthe impact of income received from PROGRESA, after controlling total house-hold resources. It is key that this estimate does not simply reflect nonlinearitiesin the effect of income on allocations; thus, the regressions control per capitahousehold expenditure with a flexible spline (with two knots at the 25th and75th percentiles of per capita household expenditure). The models also includedetailed sociodemographic controls, m, for two reasons. First, the size of thePROGRESA cash transfer depends on the age and gender composition of thehousehold, and, second, individual needsand therefore spending patternsvary with age and gender. The models also control age and education of thehead and spouse. Spending and investments into the future will also vary withcommunity-specific characteristics such as prices (including wages), levels andquality of infrastructure, ecology of the area, and the climate. Moreover, com-munities may differ in the effectiveness of implementing the program as wellas labor demand (which affects the opportunity costs of young adults attendingschool). To the extent that these effects are fixed during the study period andhave a linear and additive impact on resource allocations, they are swept out

18 On average, monthly household earnings per capita are 202 pesos in treatment households and200 in controls. The difference is 1.9 pesos, and its standard error is 1.7 pesos. This does not ruleout responses in labor supply and productivity of individuals within the household that exactlycompensate for one another and leave household earnings unchanged. We see no evidence ofdifferences in labor supply between treatment and control households. This suggests that the promiseof PROGRESA income in the future had the same impact on the expected present discountedvalue of lifetime wealth of treatment and control households. The reason may be that controlhouseholds were expecting payments (which began 2 years after the treatments) or that bothtreatment and control households expected the program to be short-lived. Net private transfers areslightly below 1 peso per capita per month in treatment households and slightly above 1 peso percapita in control households. The difference (0.05 peso) has a standard error of 0.2.


522

TABLE 4PROGRESA INCOME, OWNERSHIP OF LIVESTOCK, NUTRIENT INTAKE, AND BUDGET SHARES: MARGINAL

EFFECT OF PROGRESA INCOME (IN 000 PESOS) AFTER CONTROLLING TOTAL HOUSEHOLD RESOURCES

Dependent Variable

All Households

(Includes

Nonpoor)

(1)

Treatment

and Control

Households

(2)

Treatment

Households

Only

(3)

Treatment

Households:

In All Waves

(4)

Treatment

Households:

Always in

School

(5)

A. Livestock ownership:

Linear probability

model (probability

of having):

Chickens and turkeys 5.57 3.93 3.60 5.41 1.84

(3.12) (3.30) (3.49) (3.64) (5.37)

Pigs 13.36 13.26 14.84 14.69 6.58

(3.03) (3.16) (3.37) (3.77) (4.33)

Cows .92 1.62 2.13 .53 1.54

(2.42) (2.45) (2.50) (2.75) (3.01)

Horses and donkeys 2.51 .67 1.05 .55 1.87

(3.02) (3.11) (3.19) (3.25) (4.03)

Negative binomial

model (number of):

Chickens and turkeys .32 .38 .30 .29 .20

(.06) (.06) (.07) (.07) (.10)

Pigs .33 .33 .37 .38 .18

(.08) (.08) (.09) (.10) (.13)

Cows .21 .18 .22 .03 .14

(.13) (.12) (.15) (.16) (.23)

Horses and donkeys .08 .01 .05 .01 .05

(.09) (.09) (.10) (.11) (.15)

B. Nutrient intake:

Ln(per capita calories) .13 .15 .14 .16 .12

(.03) (.03) (.03) (.03) (.04)

Protein per calorie .15 .17 .17 .19 .15

(.03) (.03) (.03) (.04) (.05)

C. Budget shares on food:

Food 5.97 6.53 8.01 8.64 8.21

(.85) (.89) (.95) (1.04) (1.38)

Vegetables 1.51 1.77 1.95 1.74 2.09

(.33) (.35) (.39) (.43) (.55)

Fruits .05 .13 .20 .24 .24

(.09) (.09) (.10) (.10) (.13)

Tortillas and beans 2.66 2.78 3.32 3.37 3.97

(.68) (.72) (.77) (.84) (1.14)

Meat 1.88 2.14 1.71 1.53 1.76

(.57) (.59) (.64) (.72) (.92)

D. Other budget shares:

Education 2.28 2.84 3.43 3.64 2.75

(.35) (.35) (.37) (.42) (.62)



TABLE 4 (Continued )

Dependent Variable

All Households

(Includes

Nonpoor)

(1)

Treatment

and Control

Households

(2)

Treatment

Households

Only

(3)

Treatment

Households:

In All Waves

(4)

Treatment

Households:

Always in

School

(5)

Boys clothing 2.96 2.99 3.15 3.05 2.78

(.20) (.21) (.22) (.23) (.33)

Girls clothing 3.14 3.13 3.36 3.51 3.22

(.22) (.23) (.25) (.28) (.34)

Observations 31,732 22,882 14,413 11,426 6,677

Note. Column 1 includes all couples. Column 2 includes all eligible treatment and control households.Column 3 includes only treatment households. Column 4 includes treatment households that wereinterviewed in all three waves of the survey. Column 5 restricts attention to treatment households in allwaves of the survey and all of whose children were always in school in all waves. Regressions also includecommunity fixed effects, logarithm of per capita household expenditure (in a spline with knots at the25th and 75th percentiles), logarithm of household size, and number of males and females 05, 611,1225, 2645, and 45 years of age, with older females excluded; education and age of head andspouse; indicators for whether household has indoor water, electricity, concrete walls, and a concreteroof; and survey wave fixed effects (to capture time and season effects). Robust standard errors withclustering at the household level are reported below regression coefficients.

by the inclusion of a community fixed effect in the model. Variation acrossseasons and over time is captured by survey round fixed effects. All standarderrors are based on the infinitesimal jackknife and allow correlations amongunobservables at the household level. They are robust to arbitrary forms ofheteroskedasticity.

We investigate whether, when total household resources are controlled,PROGRESA income is related to investments in livestock, diet, and spendingon children. Panel A of table 4 reports the marginal effect of PROGRESAincome on the probability the household owns any animal in each category(the extensive margin) and on the number owned (intensive margin).19 Allhouseholds headed by a couple are included in the models in column 1. Thisincludes households that are poor (and thus eligible for PROGRESA) andthose that are not. The models include a control for whether the householdis a PROGRESA beneficiary, which absorbs any treatment fixed effects. Whentotal household resources are held constant, as PROGRESA income increasesand thus the share of income in the hands of women rises, so does the prob-ability of owning small livestock (chickens, turkeys, and pigs), as does the

19 Linear probability estimates are reported for the extensive margin and fixed-effects negativebinomial estimates for the intensive margin. (Restricting the latter models to assuming that theprocess is distributed as a Poisson is rejected.)



number of these small animals that are owned. (The effect on ownershipprobabilities is significant at the 5% size of the test for pigs and only the10% size of the test for chickens and turkeys.) The effects on larger animalsare both smaller in magnitude and not significant.

In case this reflects a nonlinear impact of income across its distribution,attention is restricted to poor households in column 2. Randomization wasconducted at the community level, and so the direct effect of treatment statusis absorbed by the community fixed effect. Since total household resources arecontrolled, the effect of PROGRESA income is interpreted as the effect ofincreasing the share of income under the control of women. The estimatedeffect remains significant for ownership of pigs and for the number of poultryand pigs.

Column 3 includes only treatment households and exploits the matchedadministrative records, which report the value of PROGRESA benefits paidto each household every month. In this column, it is variation in the timingof the payment that identifies the impact of income paid to women; and sincethat variation primarily reflects problems in the administration of payments,it is largely random. (Variation due to demographic characteristics of thehousehold is absorbed by the detailed demographic controls in the regression.)20

The estimated effect of PROGRESA income remains large and significant forownership of pigs and for the number of pigs and poultry.

Although PROGRESA benefits are very generous, take-up of the programis not universal. About 10% of eligible households in treatment communitiesdo not receive any PROGRESA income during the study period.21 The keycharacteristic that distinguishes eligibles who participate from those who donot is the presence (and number) of young children (aged 05). Young childrenare required to attend health clinics far more frequently than older childrenevery month among those aged under 24 months. Moreover, education benefitsare paid only for children who have passed the first three grades of primaryschool. This suggests that households with young children are more likely toview the program as providing insufficient benefits to be worth the costs ofparticipation. However, it is possible that participation is correlated withpreprogram power (or control over resources within the household), whichwould contaminate our interpretation of the results. When households withone or more children aged 5 or under are excluded, participation in the programincreases to 97% of eligibles, and sociodemographic characteristics are not

20 These are the logarithm of household size; the number of males aged 05, 611, 1225, 2645,and 1 45; and the number of females in each by the oldest age group.21 This cannot be attributed to recall error since PROGRESA income data are drawn from ad-ministrative records of actual payments.



significant predictors of participation.22 By excluding treatment householdswith young children from the sample, we can sidestep potential contaminationbecause of the participation decision. When the sample is thus restricted, theeffect of PROGRESA income on poultry and pigs does not change.23

Eligible households that did not participate in PROGRESA were also morelikely to refuse to participate in the second and third rounds of interviews.When we restrict attention to the balanced panel of households that wereinterviewed in all three follow-up surveys, program participation rates are98%, and take-up is not correlated with any of the sociodemographic char-acteristics in the models. This provides an alternative sample for assessing therobustness of our results to potential contamination due to nonparticipation.The results, reported in column 4, are very similar to those for the sample ofall treatment households. The estimated effect of PROGRESA income remainssignificant and large for pigs and for the number of poultry. Thus, part of thePROGRESA income is apparently invested, and since small livestock aretypically the domain of women in rural Mexico, the investment instrumentsappear to be those that are under the control of women. This is suggestivethat the impact of PROGRESA benefits on household resources and onwomens power within the household may be long-lived. Moreover, pigs andpoultry are often consumed by household members, and so these investmentsare likely to also contribute to improving the nutritional status of householdmembers in the future.

It would be natural to examine the effect of PROGRESA income on childanthropometry. Although such data were collected for a subsample of children,the data are not publicly available. They have been used by Behrman andHoddinot (2000), who examine the impact of participation in PROGRESAon nutritional status. They report that in treatment communities, child growthis higher and the incidence of stunting among children aged 1236 monthsis lower.24 Instead, we examine nutrient intake, in panel B. When total house-

22 The participation regressions include age- and gender-specific numbers of household members,education, and age of the head and spouse. For the sample used in col. 3, the F-statistic for jointsignificance of the demographic characteristics is 11.89 (p-value p 0.00), and the t-statistic onthe number of young children is 5.76. When households with young children are excluded fromthe sample, the F-statistic is no longer significant ( , p-value p 0.37) and no covariate isFp 1.1individually significant.23 The estimated marginal effects of PROGRESA income are not significantly different from thosein col. 3 for all outcomes in the table.24 The PROGRESA intervention involves nutrition education, which may be the proximate de-terminant of the shift toward a higher-quality diet. Since the result persists in the analyses thatare restricted to only PROGRESA households, all of which receive both the nutrition educationand income, and since the effect operates through the differential effect of PROGRESA income,



hold resources are controlled, PROGRESA income is associated with a declinein calories per capita, particularly among poor households, as well as an increasein diet quality as indicated by protein per calorie. The effects are significantacross all samples in the table. Panel C of the table focuses on the share ofthe budget spent on food.25 PROGRESA income has a negative effect on theshare of the budget spent on food, which is significant when the sample isrestricted to poor households. This reflects reduced budget shares allocated tostaples (tortillas and beans) as well as vegetables and an offsetting increase inthe share of the budget spent on meat and, thereby, improvements in dietquality.26

Panel D reports the effect of PROGRESA income on budget shares spenton goods that directly benefit children: education and childrens clothing.27

As reported by Attanasio and Lechene (2002), with resources held constant,as PROGRESA income increases, so does the share of the budget spent ongirls and boys clothing. We also find a significant and positive effect ofPROGRESA income on education. However, PROGRESA resulted in highersecondary school enrollment rates, which could explain the higher budgetshares on education and childrens clothing (Skoufias and Parker 2001; Schultz2004). To assess whether this explains the marginal effect of PROGRESA,attention is restricted to those households in which all age-eligible childrenwere enrolled in school at baseline (before the program started) and in allwaves of the survey. Results are in column 5, with the balanced panel estimatesin column 4 providing the appropriate comparison. About one-quarter andone-tenth of the marginal effect of PROGRESA income on education and onchildrens clothing, respectively, can be explained by additional enrollment.28

relative to all other income, it seems unlikely that the effect can be attributed to the nutritioneducation component of the intervention alone.25 The specification of the Engel curves in terms of budget shares has several advantages. First, itis key for our tests that nonlinearities in the demand function are captured: the share specificationperforms well in this respect and amounts to demand curves in which all covariates are interactedwith total household resources. Second, expenditure distributions are asymmetric (which suggestsusing logs) and include zeros (which makes logs unattractive); the distributions of shares are closeto symmetric, and the inclusion of zeros poses no problems in estimation. Third, the specificationhighlights how PROGRESA income is shared among goods.26 Since total expenditure is higher among treatment households, a higher meat share implieshigher spending on meat as PROGRESA income rises. In contrast, per capita expenditure on food,staples, and vegetables is not related to PROGRESA income.27 The budget shares are not exhaustive; the marginal effect of PROGRESA income on shares ofother commodity subgroups such as health, personal care, household semidurables, and entertain-ment is not significantly different from zero.28 Since treatment households that responded to the program rules by enrolling children in schoolare excluded from this subsample, the effect of PROGRESA income should decline. Nonetheless,even within the restricted subsample, education spending is significantly higher as PROGRESA



The marginal effect of PROGRESA income is positive and significant anddoes not differ significantly by gender. Moreover, the estimated effects are verysimilar across samples, indicating that the regressions do a good job of cap-turing nonlinearities in the effects of income (cols. 2 and 3) and that spendingon childrens clothing is not elevated because the children are attending school(cols. 4 and 5).

Assessment of robustness of results. The regression evidence in table 4 indicatesthat receipt of PROGRESA income is associated with shifting resources toinvestment in small livestock and in improved nutrition (more protein percalorie of intake) and in favor of child goods (childrens clothing and possiblyeducation). This is true in the entire sample and in samples that are restrictedto only poor households, to households that received the PROGRESA benefit,and to households whose age-eligible children were always in school duringthe study period.

However, in order to interpret the evidence as indicative of how resourcesunder the control of women might be spent, and as a test of the unitarymodel, it is important to assess whether alternative interpretations are con-sistent with the data. These issues are explored in table 5. The sample isrestricted to households that received the PROGRESA benefit, and the es-timates for that sample (table 4, col. 3) are repeated in column 1 of table 5.

A key assumption underlying the testing strategy is that the source ofincome has no impact on how it is invested, and PROGRESA income isdistinguished from other sources of income because PROGRESA benefits areplaced in the hands of women. An alternative approach would contrast theeffect of womens earnings and mens earnings on investment decisions. Veryfew women in the sample report any income other than PROGRESA. Only7% of women in couple households report any income, and the vast majorityof those women report only labor earnings. How a husband and wife allocatetheir time is properly treated as an integral part of the decision process un-derlying resource allocation within the household, and so comparisons of theeffects of male and female earnings on investment patterns are difficult tointerpret. The decision to invest more in small livestock may well be a con-sequence of PROGRESA shifting the household members opportunity costtoward more home-rearing activities and less to time (of women) devoted tomarket labor, and not because women have preferences different from thoseof other household members. In recognizing that time allocation is an integral

income rises. The reason may be that these households are forward looking and wish to maximizethe income they will receive from PROGRESA by increasing the chances their children will beenrolled in school throughout elementary school and the first 3 years of secondary school.


528

TAB

LE5

MA

RG

INA

LE

FFE

CT

OF

PR

OG

RE

SAIN

CO

ME

(IN00

0P

ESO

S)O

NO

WN

ER

SHIP

OF

LIV

EST

OC

K,

NU

TRIE

NT

INTA

KE

,A

ND

BU

DG

ET

SHA

RE

S:SP

EC

IFIC

ATI

ON

TEST

SA

ND

EV

ALU

ATI

ON

OF

ALT

ER

NA

TIV

EIN

TER

PR

ETA

TIO

NS

(SA

MP

LEIN

CLU

DE

SB

EN

EFI

CIA

RYH

OU

SEH

OLD

SO

NLY

)

Trea

tmen

tH

ous

eho

lds:

PR

OG

RE

SAIn

com

e(1

)

Leve

land

Vari

ance

of

PR

OG

RE

SAan

dH

ous

eho

ldIn

com

eE

xpec

ted

Ben

efit:

E(P

RO

GR

ESA

Inco

me)

(3)

Act

uala

ndE

xpec

ted

Ben

efit

Ho

useh

old

Fixe

dE

ffec

ts:

PR

OG

RE

SAIn

com

e(5

)

Ho

useh

old

Hea

dIs

PR

OG

RE

SAIn

com

e(2

a)

Var(

PR

OG

RE

SAIn

com

e)(2

b)

Var(

Ho

useh

old

Inco

me)

(2c)

PR

OG

RE

SAIn

com

e(4

a)

E(P

RO

GR

ESA

Inco

me)

(4b

)

Sing

leFe

mal

e(6

a)

Sing

leM

ale

(6b

)

A.

Live

sto

cko

wne

rshi

p:

Line

arp

rob

abili

tym

od

el(p

rob

abili

tyo

fha

ving

):C

hick

ens

and

turk

eys

3.60

4.45

.81

.04

3.05

6.99

2.72

.2

1.0

1.0

3(3

.49)

(3.5

8)(.5

7)(.1

2)(2

.61)

(3.1

1)(2

.62)

(5.7

6)(.0

2)(.0

4)P

igs

14.8

412

.43

1.00

.17

2.82

15.5

42.

094.

54

.01

.04

(3.3

7)(3

.66)

(.58)

(.13)

(2.3

3)(3

.30)

(2.3

5)(5

.85)

(.02)

(.04)

Neg

ativ

eb

ino

mia

lm

od

el(n

umb

ero

f):

Chi

cken

san

dtu

rkey

s.3

0.2

5.0

3

.001

.16

.33

.13

.33

.000

0.0

000

(.07)

(.07)

(.01)

(.008

)(.0

6)(.0

6)(.0

6)(.0

8)(.0

0002

)(.0

010)

Pig

s.3

7.3

1.0

4.0

2.0

9.4

2.0

6.3

0.0

000

.002

0(.0

9)(.1

0)(.0

2)(.0

06)

(.08)

(.09)

(.08)

(.11)

(.001

0)(.0

020)

B.

Nut

rien

tin

take

:Ln

(per

cap

ita

calo

ries

)

.14

.1

9.0

2.0

01.0

3

.15

.05

.2

0.0

001

.000

2(.0

3)(.0

4)(.0

05)

(.002

)(.0

3)(.0

3)(.0

3)(.0

7)(.0

002)

(.000

4)


529

Pro

tein

per

calo

rie

.17

.23

.0

2

.003

.0

1.1

7

.03

.21

.000

3.0

001

(.03)

(.04)

(.01)

(.002

)(.0

3)(.0

4)(.0

3)(.0

8)(.0

002)

(.001

0)C

.B

udg

etsh

ares

on

foo

d:

Foo

d

8.01

7.

56

.16

.14

4.

55

6.99

3.

00

.65

.002

.0

2(.9

5)(1

.13)

(.18)

(.04)

(.77)

(.99)

(.80)

(1.9

9)(.0

08)

(.03)

Veg

etab

les

1.

95

3.25

.33

.0

1

1.02

1.

75

.63

.5

5

.004

.0

1(.3

9)(.5

0)(.0

8)(.0

2)(.3

2)(.4

0)(.3

3)(.8

8)(.0

03)

(.01)

Frui

ts

.20

.1

7

.01

.00

.1

1

.17

.0

7

.24

.000

4

.004

(.10)

(.13)

(.02)

(.000

2)(.0

7)(.1

1)(.0

8)(.2

1)(.0

01)

(.004

)To

rtill

asan

db

eans

2.

66

3.69

.10

.22

1.

29

3.13

.6

0

2.44

.000

1.0

1(.6

8)(.9

6)(.1

5)(.0

6)(.6

3)(.8

0)(.6

5)(1

.74)

(.008

)(.0

2)M

eat

1.88

3.24

.4

4

.04

.1

51.

91

.58

2.28

.006

.0

1(.5

7)(.7

4)(.1

1)(.0

3)(.5

1)(.6

5)(.5

2)(1

.31)

(.01)

(.02)

D.

Oth

erb

udg

etsh

ares

:E

duc

atio

n3.

432.

14.3

8.0

02.

462.

801.

84.2

2

.001

.0

1(.3

7)(.4

5)(.0

8)(.0

1)(.2

9)(.3

8)(.3

0)(.8

0)(.0

02)

(.02)

Bo

ys

clo

thin

g3.

152.

9.0

7

.04

1.48

2.86

.85

1.11

.001

.01

(.22)

(.26)

(.04)

(.01)

(.17)

(.23)

(.18)

(.43)

(.002

)(.0

1)G

irls

cl

oth

ing

3.36

3.19

.05

.0

11.

423.

11.7

32.

11.0

01.0

1(.2

5)(.2

9)(.0

4)(.0

1)(.1

7)(.2

6)(.1

8)(.4

8)(.0

02)

(.01)

No

te.

Co

lum

n1

rep

eats

col.

3o

fta

ble

4fo

rre

fere

nce.

Co

lum

ns2

sho

wth

em

arg

inal

effe

cto

fP

RO

GR

ESA

ben

efit

(co

l.2a

)af

ter

cont

rolli

ngva

rian

ceo

fth

eb

enefi

t(c

ol.

2b)

and

vari

ance

of

tota

lho

useh

old

inco

me

(co

l.2c

).C

olu

mn

3re

po

rts

the

effe

cto

fex

pec

ted

PR

OG

RE

SAb

enefi

t.C

olu

mns

4re

po

rtth

eef

fect

of

PR

OG

RE

SAb

enefi

t(c

ol.

4a)

cont

rolli

ngex

pec

ted

ben

efit

(co

l.4b

).C

olu

mn

5re

po

rts

the

effe

cto

fP

RO

GR

ESA

ben

efit

afte

rco

ntro

lling

hous

eho

ldfix

edef

fect

s.C

olu

mns

15

use

the

sam

ple

of

ben

efici

ary

hous

eho

lds

head

edb

ya

coup

le.

Co

lum

ns6

rep

ort

the

effe

cto

fP

RO

GR

ESA

ben

efit

for

two

diff

eren

tsa

mp

les:

hous

eho

lds

head

edb

ya

sing

lefe

mal

e(c

ol.

6a)

and

hous

eho

lds

head

edb

ya

sing

lem

ale

(co

l.6b

);sa

mp

lesi

zes

are

658

and

244,

resp

ecti

vely

.Se

eth

eno

teto

tab

le4.



part of household decision making, studies have used income from nonlaborsources. If we put aside the strong assumptions needed to interpret thosemodels, less than 1% of women in this sample report any nonlabor income,and so this is not a practical approach with these data. Thus, treating PRO-GRESA income as equal to female nonlabor income is not an unreasonableapproximation in our context. Clearly a key advantage of PROGRESA incomein this study is that it can legitimately be treated as an exogenous increasein income that is placed in the hands of women in the (randomly assigned)treatment communities.

Nevertheless, PROGRESA income may differ from other income sourcesbecause it is a government transfer and is not subject to, say, the vagaries ofthe weather, as would be the case for income from agriculture, for example.More generally, how income is invested may differ depending on the pre-dictability of the income. To address this question, the three follow-up surveysare exploited to calculate the household-specific variance of PROGRESA in-come and also of total household income. They are included with PROGRESAincome in columns 2a2c of table 5.

The positive effects of PROGRESA income on the probability of owningpigs as well as the number of pigs and poultry owned are robust to theinclusion of variances, and the variance of the benefit also has a positive effecton these investments (suggesting that savings may come out of transitoryincome). The inclusion of income variances in the model increases the positiveeffect of PROGRESA income on the quality of nutrient intake linked to anincrease in meat spending (although higher variances in benefits do depressmeat shares and increase shares of vegetables) but has little impact on theestimated effects of PROGRESA income on other foods and on childrensclothing. Education shares rise as the variance of the PROGRESA benefitsrises: this likely reflects reverse causality since a higher variance in benefitsimplies a response to the incentives in the program (by changing attendanceat schools or health clinics). Conditional on that variance, the level of PRO-GRESA continues to be positively associated with education shares, suggestingthat the effect is not entirely due to reverse causality. We conclude that therejection of the unitary model is not driven by differences in the variances ofPROGRESA income relative to other income.

The analyses thus far have relied on actual payments of benefits to households(from administrative records). These amounts differ from expected paymentson the basis of the program rules because of errors in administration and delaysin payments. It is possible that households use expected payments when mak-ing allocation decisions, in which case actual payments will be an error-ridden



proxy for the relevant construct; if the reverse is true, expected payments willbe noisy proxies.

Estimates of the effects of the expected benefit, reported in column 3 oftable 5, are smaller (in absolute value) than the effects of the actual benefit(col. 1 of the table). This suggests that decisions are based on income receivedrather than on expected income. To probe this further, we estimated themodel and its reverse,Actual p a b Expected Expected p a i 0 0 i I i 1

. If the two measures are identical, then b will be one and ab Actual u1 i iwill be zero, whereas deviations from those values indicate that the independentvariable is measured with (classical) error. Since andb p .45 a p 0.10 0whereas and , it appears that expected payments are error-b p 1.1 a p 0.071 1ridden proxies for actual payments.

Models that include both expected and actual benefits are reported in col-umns 4a and 4b of table 5. The effects of actual benefits are very similar tothose in column 1 (and none of the differences is significant). Expected benefitshave a significant impact only on the number of poultry owned, food shares,education, and childrens clothing. In all cases, these effects are smaller thanthe impact of actual benefits. All the evidence points to household decisionsbeing based on the benefits at the time they are paid to the women and noton the potential income from the program. If households are able to borrowagainst future PROGRESA income and there is no uncertainty about whetherPROGRESA benefits will be paid, expected benefits should affect decisionmaking. The evidence is suggestive that PROGRESA beneficiaries either facebinding liquidity constraints or are uncertain about payments or both.

In a further exploitation of the longitudinal dimension of the data, themodels include a household fixed effect that will sweep out all characteristicsof households that do not change during the 18-month study period. Thisincludes household permanent income (and the expected PROGRESA benefitas long as that expectation is fixed) along with any behavioral response to theinitiation of the program (such as a change in time allocation of householdmembers, interhousehold transfers, or a change in choice of crop or technology).The fixed effect will also absorb household-specific differences in measurementerror of household resources and PROGRESA benefits. Results are reportedin column 5 of table 5. The estimates measure the effect of changes in PRO-GRESA benefits on household resource allocations. In the absence of liquidityconstraints, the effects should be zero. While the effects are smaller than thosein column A, they are not zero. PROGRESA income continues to increasethe number of small livestock (poultry and pigs) that are owned, reduce caloriesper capita, improve diet quality, and increase the share of the budget spenton meat, as well as childrens clothing. The evidence indicates that additional



income in the hands of women results in shifting resources toward investmentsin small livestock, nutrition, and children.

The possibility remains that the results presented thus far have nothing todo with giving money to women, but are, instead, related to some otherdimension of the program, and its impact is directly related to the share ofhousehold income that is received from PROGRESA. For example, one of themessages of PROGRESA is that parents should invest in a better diet and intheir children. Participation in the program may affect the returns to savingand investing in children or income from the program may directly affectbehaviors related to those choices. These effects should be apparent not justfor couples but also for single-headed households. Columns 6a and 6b of thetable presents results for households headed by single women (in 6a) and singlemen (in 6b). The evidence is unambiguous: PROGRESA income has no impacton any of the budget allocations or on investments in livestock among single-headed households. When total household resources are controlled, the mar-ginal effects of PROGRESA are both substantively very small and not sig-nificantly different from zero, indicating that in single-headed householdsPROGRESA income is treated as any other income source. We conclude thatthe results regarding the impact that PROGRESA income given to womenin couple households has on resource allocations cannot be attributed to un-observed heterogeneity that is associated with the PROGRESA program.

In sum, the evidence indicates that PROGRESA payments to women aredirected toward investments that are likely to remain under their control.Under the assumption that women are more patient than men, the evidencecan be interpreted as indicating that the PROGRESA payment is associatedwith an increase in the womans power within the household, and she has agreater say in resource allocations within the household. The PROGRESAsurvey does not contain direct evidence on preferences of individuals, and sowe turn to data from the Mexican Family Life Survey.

VI. Direct Evidence on Intertemporal PreferencesThe MxFLS is a broad-purpose, nationally representative, longitudinal surveyof individuals, households, and their families (Rubalcava and Teruel 2007)that collects extensive information on socioeconomic and demographic char-acteristics of respondents, including participation in public programs. Thebaseline, conducted in 2002, interviewed respondents in over 8,400 house-holds, and more than 90% of those respondents were reinterviewed in thesecond wave in 2005.

Key for this research is that in the second wave of MxFLS, every householdmember aged 15 and older completed a module that seeks to elicit intertem-



TABLE 6INTERTEMPORAL PREFERENCES OF ADULT MALES AND FEMALES IN RURAL MEXICO: DISTRIBUTION OF

CHOICES OF HYPOTHETICAL PAYMENT NOW OR IN 3 YEARS TIME BY GENDER OF RESPONDENT

Win $10,000 in the Lottery and CollectFemales

(1)Males

(2)Difference

(3)

1. $10,000 today (most impatient) .52 .53 .01(.01) (.01) (.01)

2. $40,000 in 3 years .12 .17 .04(.01) (.01) (.01)

3. $20,000 in 3 years .12 .12 0(.01) (.01) (.01)

4. $15,000 in 3 years .11 .09 .02(.01) (.01) (.01)

5. $12,000 in 3 years (most patient) .13 .09 .04(.01) (.01) (.01)

Note. Results from hypothetical questions asked of rural respondents in the second wave of theMxFLS. Number of observations is 5,230. Standard errors are in parentheses.

poral preferences. The respondent is asked to consider a hypothetical situationin which he or she wins the lottery and can be paid now or paid a larger sumin the future. The first battery of questions asks about payment now relativeto 1 month in the future. The second battery asks about payment now relativeto 3 years hence. In each battery, the respondent is given a series of binarychoices of payment now or a higher payment later with increasing impliedinterest rates. The frequency distribution of choices for the longer-term batteryis displayed in table 6 for females (col. 1), males (col. 2), and the differencebetween females and males (col. 3). Slightly over half of the respondents areimpatient in the sense that they always preferred to be paid now relative tolater. About 10% are patient (they prefer $12,000 in 3 years to $10,000 now).Females are generally more patient than males.

The evidence is summarized in a regression framework in table 7, whichreports estimates from an ordered probit regression in which the dependentvariable ranges between 1 (most impatient) and 5 (most patient) as in table6. Robust standard errors, which allow for clustering at the household level,are reported below regression coefficients. We report the difference betweenfemales and males for the longer-term (row 1) and shorter-term (row 2) choices.The regression results indicate that females are significantly more patient thanmales. This is true when the regressions include no controls (col. 1), whenthe estimates are adjusted for age and education (col. 2), and also when theyare adjusted for household composition and resources (col. 3). About one-quarter of the sample respondents live in PROGRESA beneficiary households.In order to parallel the evidence based on the PROGRESA data, analysis isrestricted to adults living in beneficiary households in column 4. In thesehouseholds females are significantly more patient than males, and the difference



TABLE 7INTERTEMPORAL PREFERENCES OF ADULT MALES AND FEMALES IN RURAL MEXICO: ORDERED PROBIT

ESTIMATES OF DIFFERENCES IN TIME PREFERENCES OF FEMALES RELATIVE TO MALES

(CHOICE OF PAYMENT NOW VS. 3 YEARS AND NOW VS. 1 MONTH)

Unconditional(1)

Controlling Ageand Schooling

(2)

Plus ControlsHousehold

Compositionand Resources

(3)

PROGRESABeneficiaries

(4)

Longer term (3 years vs. now) .11 .10 .10 .13(.03) (.03) (.03) (.05)

Shorter term (1 month vs. now) .12 .11 .10 .13(.03) (.03) (.03) (.05)

Observations 5,230 5,230 5,230 2,315

Note. The regression in col. 1 includes no additional controls. Column 2 controls the respondents ageand schooling. Column 3 also controls household composition (logarithm of household size and numberof males and females 05, 611, 1225, 2645, and 45 years of age) and the logarithm of per capitahousehold expenditure (specified as a spline with two knots at the 25th percentile and at the 75thpercentile). Column 4 repeats the model in col. 3 restricting the sample to individuals living in householdsreporting receipt of PROGRESA. Robust standard errors that allow for clustering at the household levelare reported below regression coefficients.

is slightly larger than in other rural households. The direct measures provideunambiguous evidence that there is heterogeneity in preferences between malesand females, with the latter being significantly more patient when it comesto financial decisions.29

Each adult respondent was also asked to choose between some hypotheticalgambles in an effort to assess their willingness to take on financial risk. Riskaversion and time preferences are likely to be related, and so it is possiblethat the fact that females tend to be more patient than males is a reflectionof differences in attitudes toward risk. We find no differences in the expectedvalue of the riskiest gamble that female respondents are willing to take relativeto male respondents. For example, in rural areas, females are about 2% morelikely to take a more risky gamble than males, and among PROGRESArecipients, the difference is about 3%. In both cases, the difference is notsignificant (t-statistic is 1.3). We conclude that it is differences in time pref-erences of males and females that drive the decisions about how to allocatePROGRESA income.

29 The hypothetical questions follow the protocols used in behavioral economics incentivized tasksalbeit with a coarser grid. A small subsample of respondents completed incentivized tasks withreal stakes, and similar patterns emerge for both the incentivized tasks and hypothetical questions.In both cases, females are more patient than males and younger respondents are more impatientthan older respondents. Preliminary evidence suggests that the hypothetical questions and incen-tivized tasks provide a consistent picture of characteristics that are correlated with preferences inthe rural Mexican population. Note that hypothetical questions have some advantages: contami-nation due to credibility of future payments, liquidity constraints, and the budget constraint ofthe experiment are not relevant. See Eckel et al. (2006) for more detail.



VII. ConclusionsPROGRESA benefits, which were paid to women, increased total householdincome by around one-quarter among those rural Mexicans who received thebenefit. The impact of additional income in the hands of women is examinedby exploiting the fact that otherwise identical communities were randomlyassigned to be treatment communities, in which eligible households receivedthe benefit at the beginning of the study period, or to control communities,in which eligible households would receive the benefit after the study period.

In households that are headed by couples, relative to other household income,PROGRESA income is spent on small livestock, higher-quality nutrient in-take, and childrens clothing. In households headed by single females or singlemales, PROGRESA income is treated no differently than any other income.Qualitative evidence from interviews conducted with PROGRESA householdsindicates that PROGRESA income was perceived as being under the controlof women: Now we dont demand, every moment, give me for shoes, give mefor that. Now we take the money from PROGRESA and we buy from thatmoney. Now we dont bother them [their husbands] so much (Adato et al.2000, 63).

Moreover, direct evidence on preferences indicates that women in ruralMexico are more patient than men. Taken together, the evidence suggests thatPROGRESA benefits increase the power of women, who are better able toassert their preferences and allocate more resources within households towardinvestments. These investments include small livestock (over which they havesome control) and their children. This is an important result. First, it suggeststhat empowering women is likely to be associated with elevated levels ofsavings and investments, which will, in turn, likely contribute to futuregrowth. Second, the results suggest one mechanism for the empirical findingthat mothers allocate more resources to their children than fathers: womenare inclined to invest more in the future, their children.

ReferencesAdato, Michelle, Benedicte de la Brie`re, Dubravka Mindek, and Agnes Quisumbing.

2000. Final Report: The Impact of PROGRESA on Womens Status and Intra-household Relations. Report, International Food Policy Research Institute, Wash-ington, DC.

Andreoni, James, and Lise Vesterlund. 2001. Which Is the Fair Sex? Gender Dif-ferences in Altruism. Quarterly Journal of Economics 116, no. 1:293312.

Arizpe, Lourdes, Carlota Botey, and Carmen D. Deere. 1986. Mexican AgriculturalDevelopment Policy and Its Impact on Rural Women. In Rural Women and StatePolicy: Feminist Perspectives on Latin American Agricultural Development, ed. Magdalena



Leon, 6783. Series in Political Economy and Economic Development in LatinAmerica. Boulder, CO: Westview.

Arriaga-Jordan, Carlos M., and R. A Pearson. 2004. The Contribution of Livestockto Smallholder Livelihoods: The Situation in Mexico. In Responding to the LivestockRevolution: The Role of Globalisation

investments, tp and tranfers made to women

Documents