77400 the world bank economic review, vol. 17, no. 1 27–50 Public Policy and Extended Families: Evidence from Pensions in South Africa Marianne Bertrand, Sendhil Mullainathan, and Douglas Miller How are resources allocated within extended families in developing economies? This question is investigated using a unique social experiment: the South African pension program. Under that program the elderly receive a cash transfer equal to roughly twice the per capita income of Africans in South Africa. The study examines how this trans- fer affects the labor supply of prime-age individuals living with these elderly in extended families. It finds a sharp drop in the working hours of prime-age individuals in these households when women turn 60 years old or men turn 65, the ages at which they become eligible for pensions. It also finds that the drop in labor supply is much larger when the pensioner is a woman, suggesting an imperfect pooling of resources. The allo- cation of resources among prime-age individuals depends strongly on their absolute age and gender as well as on their relative age. The oldest son in the household reduces his working hours more than any other prime-age household member. In many developing economies large extended families often live together. Shared housing may suggest the sharing of other resources, most notably money. If such resource sharing is prevalent, social policies may produce unexpected outcomes. A transfer targeted to one demographic group may eventually find itself in the pockets of relatives living in the same house. Who in the end benefits from the transfer will depend on the sharing rules within the household. To understand how resources are transferred in extended families,1 this study investigates South Africa’s unusual old-age pension program. The program grants Marianne Bertrand is Associate Professor of Economics, University of Chicago Graduate School of Business, Center for Economic and Policy Research, and National Bureau of Economic Research. Her e-mail address is marianne.bertrand@gsb.uchicago.edu. Sendhil Mullainathan is Associate Professor of Economics, Massachusetts Institute of Technology and National Bureau of Economic Research. E-mail: mullain@mit.edu. Douglas Miller is Assistant Professor of Economics, University of California, Davis. His e-mail address is dlmiller@uc.davis.edu. The authors are grateful to two anonymous referees, Abhijit Banerjee, Anne Case, Angus Deaton, Esther Duflo, Jon Gruber, Michael Kremer, Jonathan Morduch, and Jim Poterba for many helpful comments. They have also benefited from feedback from seminar participants at the MIT Public Finance Lunch, Princeton Development Workshop, Harvard-MIT Devel- opment Seminar, and the National Bureau of Economic Research Summer Institute 2000. Miller ac- knowledges financial support from the National Science Foundation’s Graduate Fellowship Program. 1. A large body literature has examined resource transfers in the close family (husband and wife or parent and young child). Lundberg and Pollak (1996) provide a survey. In the close family, one can rea- sonably assume that resource transfers take place, for example, between parents and young children. The DOI: 10.1093/wber/lhg014 © 2003 The International Bank for Reconstruction and Development / THE WORLD BANK 27 28 the world bank economic review, vol. 17, no. 1 large lump-sum cash transfers, roughly twice the average per capita income in African households, to eligible women over the age of 60 and men over the age of 65.2 The magnitude of the transfer makes it a useful experiment, permitting the tracking of the flow of money more cleanly than would more marginal changes. Does the pension money eventually reach family members other than the pensioners? If so, how much of the cash is transferred, and which family members receive most of it? These questions are addressed through an examination of the labor supply of relatives living with pensioners. This approach has two advantages. First, typical household survey data do not allow direct measuring of the transfers to each fam- ily member. The survey data used in this study are no exception. Expenditure data measure consumption at the household, not the individual level. Only a few con- sumption items are exclusive enough that they can be matched to a specific gen- der or age group.3 Leisure time, however, is a good that can easily be assigned (Chiappori 1992). Labor supply data can be used to infer (at least partly) how the pension money is allocated among the prime-age individuals in a household. Second, a labor supply response would most clearly underline the unexpected outcomes caused by family redistribution. Because the social pension targets a group that by and large is already out of the labor force and conditions mainly on an unalterable variable, age, it might be expected to have little effect on labor supply.4 On the other hand, if intrahousehold redistribution occurs, aggregate labor supply may fall as the prime-age individuals who live with pensioners re- duce their hours of work. The size of the effect will depend on the direction and strength of redistribution flows inside households. Anecdotal evidence and newspaper articles hint that the pension may well have affected relatives’ labor supply. One article mentions that “the impact of pen- sions on communities with a high rate of unemployment was huge, as multi- generation households formed a constellation around the person receiving the pension� (Ngoro 1998). Another describes a pensioner’s “five children, who also live with him in his two-bedroom flat, contribute to the family income when they can find work. But none has a full-time job� (Caelers 1998). Of course, such rare evidence on resource transfers in the extended family comes from the United States (Altonji and others 1992) and suggests that resource transfers are not large in that case. It is an open question whether such a finding generalizes to developing economies, where the extended family often lives under a com- mon roof. 2. The survey data used for this study classified people into four different racial groups: white, col- ored, Indian, and African. This study looks only at African households. In theory, the transfer program is means-tested, but this has little effect in practice for Africans whose income is quite low relative to the test. 3. Subramanian and Deaton (1991) use expenditures on adult goods, such as alcohol and tobacco, to study discrimination based on children’s gender. Browning and others (1994) use expenditures on men’s and women’s clothing to study sharing rules between couples. 4. A labor supply effect might arise because the pension increases the expected future income of the young. But this would affect all the young equally, not merely the relatives of pensioners, as the results suggest. Bertrand, Mullainathan, and Miller 29 stories do not constitute causal evidence, but they provide a backdrop for the statistical work presented here. The study uses the sharp rise in household income when an elderly member crosses the pension age threshold to identify the pension’s effect. The results suggest that the pension dramatically reduces the labor supply of the prime-age members of the household. Both hours worked and the work or not-work margin are affected. A clear discontinuity appears exactly at the age-eligibility frontier, with labor supply by the household dropping when a woman in the household reaches age 60 or a man reaches age 65. Roughly speaking, the age of the elderly does not seem to affect labor supply except at these discontinuous points. Absolute age, relative age, and gender are important determinants of resource flows in that they affect the strength of the labor supply response. Holding fam- ily composition constant, the study finds that the marginal rand of pension in- come going to a female pensioner reduces labor supply more than the marginal rand of pension income going to a male pensioner. This gender impact on the flow of resources suggests that common-preference models of the family, which view the family as maximizing one common utility function, and for which the source of the pension income should not matter, may not fit these extended fami- lies very well. The study also finds that prime-age women reduce their labor supply less than prime-age men for each marginal rand of pension money received by the elderly. Also, working hours drop more as the age of the prime-age family member in- creases. Finally, after controlling for the differential effect of the pension by gender and age, the study finds that the oldest prime-age male in a household reduces his labor supply more than do other prime-age household members. In summary, although the South African pension program was introduced as a way to improve living standards among elderly people who do not have access to a private pension, the results show that intrahousehold redistribution sub- stantially reduces the size of the transfer to that demographic group. At least part of the pension money ends up with a group that was not originally targeted: prime-age individuals that live with the pensioners I. The Old-Age Pension Program The social pension program in South Africa, which dates to the 1920s, was his- torically intended for white South Africans only.5 Disintegration of the apart- heid regime in the late 1980s and early 1990s led to pressures for more racial parity in pension eligibility and benefits. Major reforms of the pension program for African households took place after 1992, with the introduction of superior technologies in the pension delivery system (in part to improve access to remote 5. Additional information about the historical background, institutional features, and practical implementation of this program can be found in Lund (1992), Van der Berg (1994), and Case and Deaton (1998). 30 the world bank economic review, vol. 17, no. 1 areas) and the equalization of both the means-test and the pension benefit levels across racial groups. Eligibility for pension receipt is determined primarily by age: only women over the age of 60 and men over the age of 65 are eligible. In practice, though, some local authorities have been equalizing the pension eligibility age for men and women. Hence, a nontrivial share of men between 60 and 65 years of age report receiving a pension. (This fact is exploited later in the analysis of the effect of the pension.) The state social pension is means tested, with the result that most whites are excluded from the pension whereas most Africans are entitled to the maximum benefits. Case and Deaton (1998) show that 14 percent of white women and 7 percent of white men report receiving the pension, compared with 80 percent of African women and 77 percent of African men.6 The South African social pension is very generous. The maximum benefit in 1993, the year the survey data were collected, was 370 rand a month, or about half the average African household income and more than twice the median per capita income among Africans. Such large pension transfers could be expected to result in intrahousehold redistribution that leads to significant behavioral responses, such as a reduced willingness to participate in the labor force among family members not originally targeted by policymakers. II. Data and Summary Statistics The primary data set used in this article is the Integrated Household Survey of South Africa. This survey is the result of a cooperation between the World Bank and the South African Development Research Unit at the University of Cape Town.7 The survey, a random sample of 9000 households, was conducted dur- ing the second half of 1993. Means-testing for the pension is such that only a small share of elderly white women and white men report receiving any pension transfer, and the participation rate for colored and Indian South Africans, though higher, is well below African rates (Case and Deaton 1998). Moreover, the preva- lence of multigeneration households is much larger among Africans than among the other racial groups (Ardington and Lund 1994). To keep the focus on extended families, the study was restricted to three- generation households (a household containing at least a child, a parent, and a grandparent). This restriction also reduces the heterogeneity in the sample. With- out it, pension-ineligible households could also include individuals living away from their elders. Because such individuals would clearly be different from those living with their elders, this could introduce a selection bias. The restriction to 6. The means- testing formula does not take into account income from family members other than the elderly (Case and Deaton 1998). Hence, there are no direct incentives in the program design for family dissolution or migration. 7. The database used in this article can be downloaded directly from www.worldbank.org/html/ prdph/lsms. Bertrand, Mullainathan, and Miller 31 three-generation households guarantees that the age of the elderly is the only source of variation. The study looks at the labor supply of working-age individuals between 16 and 50 years old (prime age) in these multigeneration households. A conserva- tive cut-off age of 50 years is used to avoid any effect arising because people expect to get the pension themselves soon. More than a third of prime-age indi- viduals in the original sample live in three-generation households, as do a large proportion of women over age 60 and men over age 65, a fact previously noted by Case and Deaton (1998).8 The dependent variable in most of the regressions reported here is weekly working hours for prime-age individuals. For each person 16 years old or older, the survey asks: “How many hours did ______ work last week?� The working hours question relates to all forms of employment: regular wage employment (self-employed professionals), casual wage employment, self-employment in agriculture, and other forms of employment and self-employment. The analysis also sometimes uses a dummy variable for employment status as a measure of labor supply. Again, the employment status variable refers to all forms of em- ployment and not exclusively to regular employment. The study also briefly documents whether any change in employment status reflects a change in unemployment or labor force participation status. Individu- als who report not being currently employed are asked if they have been looking for work during the previous week. Answers to these two questions are used to classify people as employed, unemployed, or not in the labor force. Individuals out of the labor force are then asked why they did not look for work in the pre- vious week. Individuals out of the labor force who did not look for work be- cause they thought there were “no jobs or work available� are further classified as discouraged workers. Table 1 presents means and standard deviations of the main variables of in- terest for African individuals between the ages of 16 and 50 who live in three- generation households. Because identification of the pension impact eventually relies on the presence or not of age-eligible people in the household, these means and standard deviations are also presented separately for households with at least one age-eligible person (woman over 60, man over 65) and households without. Several interesting facts emerge from table 1. First, only 23 percent of people in the sample are employed. The employment rate is 26 percent among men and 21 percent among women. Average working hours, 6.3, are also very low. Of the remaining 77 percent of people who are not employed, 8 percent are unem- ployed and 21 percent are discouraged, leaving roughly 48 percent out of the labor force and not discouraged. The low employment rate and high discour- agement and unemployment rates among prime-age African individuals is a well- documented characteristic of South African labor markets. 8. As expected, households that contain eligible elderly but that are not three-generation house- holds are on average much smaller (a little less than four people on average) and older. 32 the world bank economic review, vol. 17, no. 1 Table 1. Descriptive Statistics, 16- to 50-Year-Old Africans in Three-Generation Households All Age-eligible Non–age-elegible households households households Variable Mean SD Mean SD Mean SD Age 27.5 9.3 27.5 8.7 27.5 9.9 Employed 0.229 0.420 0.212 0.409 0.246 0.431 Hours worked 6.32 16.37 3.21 12.51 9.45 19.00 Unemployed 0.079 0.270 0.087 0.232 0.071 0.256 Discouraged 0.211 0.408 0.232 0.422 0.191 0.393 4th grade or more 0.754 0.431 0.748 0.434 0.760 0.427 8th grade or more 0.348 0.477 0.338 0.473 0.360 0.480 Matric or more 0.130 0.336 0.128 0.335 0.132 0.338 Household size 8.81 3.62 9.13 3.88 8.50 3.30 Rural 0.683 0.465 0.707 0.455 0.660 0.474 Urban 0.166 0.372 0.152 0.359 0.180 0.384 Metro 0.151 0.358 0.141 0.348 0.161 0.367 Sick 0.065 0.246 0.073 0.261 0.056 0.230 Total income 1325 1833 1318 1246 1333 2272 Pension income 207 275 371 277 42 142 Number of eligible women 0.454 0.526 0.906 0.377 0 0 Number of eligible men 0.169 0.383 0.338 0.485 0 0 Note: Sample is composed of set of African individuals between 16 and 50 years old that live in a three-generation household. Sample size: all households, 6,326; age-eligible households, 3,169; non–age-eligible households, 3,157. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. Second, on background characteristics, the differences between eligible and ineligible households are small. For example, there are only limited differences in education or in geographical distribution across rural and urban areas. Age- eligible households appear a little bigger on average (9.1 versus 8.5).9 One notice- able difference is that prime-age individuals in eligible households report being sick more often.10 Third, on employment status and working hours, the difference between the two household types is dramatic. Raw differences in employment rates are more than 3 percentage points. The econometric work reported translates these raw differences into estimates of the effect of the pension. The analysis shows some other interesting patterns. Pension income in eligible households accounts for more than a quarter of total household income, dem- onstrating the generosity of the social pension program. The average eligible household has 0.9 eligible women and 0.34 eligible men, for a total of 1.24 eli- 9. A similar exercise performed for all prime-age individuals, not only those living in three-genera- tion households, produces dramatic differences on such variables, underlining the importance of focus- ing on three-generation households. 10. One might argue that sickness is a luxury good among these African households and that this might be looked at as an outcome of the social pension. Bertrand, Mullainathan, and Miller 33 gible members (table 1). Most of the pension income, therefore, comes through a woman. Many households have more than one pensioner. III. Basic Results The first set of regressions compares the labor supply of prime-age individuals who live with age-eligible elderly relative with those who do not and considers the effect for both men and women (table 2). Each regression includes, in addition to the pension variable, a quartic in individual age, a dummy variable for whether the individual completed eighth grade, 14 province dummy variables, 3 location dummy variables (rural, urban, and metropolitan area), a female dummy variable, house- hold size, and number of household members 0–5 years old, 6–15, 16–18, 19–21, and 22–24 years old.11 For these results and all the results that follow, standard errors are corrected to allow for correlation in outcomes within household clusters. Both working hours (columns 1–3) and employment status (columns 4–6) are used dependent variables (table 2). Basic ordinary least squares (ols) regressions of labor supply on continuous pension income (columns 1 and 4) show that more pension income significantly reduces both working hours and employment rates. The simple ols results, however, are not exploiting only the variation in pen- sion receipt that comes from the age of the elderly household members. By using information on actual pension receipt, the estimates may be biased by endog- enous takeup or eligibility. Takeup rates are high but not complete, and although the means test is low, some elderly do fail to get the pension. If those who actu- ally receive the pension are different from those who do not, the ols estimate will be biased. This possibility is addressed by examining the effect of pension eligibility (the age-eligibility criterion) rather than actual pension receipt. A similar negative labor supply response is found for households that have at least one age-eligible person compared with households that do not (columns 2 and 5). This eligibility measure cannot easily be transformed into a meaningful economic measure (such as an elasticity with respect to pension benefits). To ease economic interpretation, the continuous pension income variable (the amount of pension benefits received by a household) is instrumented with the number of age-eligible men and women in that household (columns 3 and 6). The first-stage regressions associated with columns 3 and 6 (not reported here) show that the numbers of age-eligible women and men are both very significant determinants of monthly pension income. The coefficient on number of women over age 60 and number of men over age 65 are very similar. The null hypothesis that these two coefficients are the same at standard confidence levels, and hence that the men and women have similar takeup rates, cannot be rejected. The instrumental variable (iv) coef- 11. The completion of matric (10th grade) is another important determinant of employment and unemployment probabilities among South African men and women. The results are unaffected if we use the completion of 10th grade instead of the completion of 8th grade as a control for educational attainment. 34 Table 2. Effect of Old-Age Pension Income on Working Hours and Employment Status of 16- to 50-Year-Old Africans Hours worked Employment status ols ols Pension Pension Pension Pension uptake eligibility iva uptake eligibility iv Variableb (1) (2) (3) (4) (5) (6) Pension income × 1000 –12.32 (1.18) — –17.07 (1.78) –0.053 (0.022) — –0.099 (0.035) Household eligibility — –6.401 (0.580) — — –0.043 (0.011) — dummy Female –2.552 (0.447) –2.666 (0.452) –2.629 (0.452) –0.068 (0.012) –0.069 (0.012) –0.069 (0.012) 34 Age –6.526 (3.640) –6.732 (3.709) –6.585 (3.611) –0.394 (0.090) –0.395 (0.090) –0.394 (0.090) Age2 0.407 (0.187) 0.412 (0.190) 0.404 (0.185) 0.022 (0.055) 0.022 (0.055) 0.021 (0.055) Age3 –0.010 (0.004) –0.010 (0.004) –0.10 (0.004) –0.0004 (0.0001) –0.0004 (0.0001) –0.0005 (0.0001) Age4 × 1000 0.082 (0.032) 0.081 (0.032) 0.080 (0.032) 0.0036 (0.0008) 0.0036 (0.0008) 0.0036 (0.0008) the world bank economic review, vol. 17, no. 1 8th grade or more 1.485 (0.466) 1.262 (0.468) 1.520 (0.469) 0.064 (0.012) 0.062 (0.012) 0.064 (0.012) R2 0.126 0.123 — 0.192 0.193 — Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size in all regressions is 6,326. aPension income is instrumented with the number of age-eligible women and age-eligible men in the household. bOther covariates included in regression are 14 province indicators, 3 location indicators (urban, rural, and metro), household size, number of household members ages 0–5, 6–15, 16–18, 19–21, and 22–24. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. Bertrand, Mullainathan, and Miller 35 ficients on pension receipt in columns 3 and 6 are even more strongly negative than the ols coefficients in columns 1 and 4. Each extra 100 rand of pension in- come reduces weekly labor supply of prime-age individuals by about 1.7 hours.12 How large are these effects? For simplicity, assume that the pension is split equally across all prime-age household members.13 Because there are 4.7 prime- age people in the average household, the coefficient of –17.07 suggests that a 1000 rand change in individual income reduces hours worked by –17.07 times 4.7 (table 2). Average individual income (computed as household income divided by number of prime-age people in the household) is 272 rand. Average hours, conditional on working, equal 41.4.14 Scaling by these gives an elasticity of hours to income of –17.07 times 4.7 times [(0.272) / (41.4)] equals –.53. Similarly, the elasticity for employment is computed as –0.099 times 4.7 times [(0.272) / (0.229)] equals –0.55. These elasticities are large if viewed as pure income effects (see Imbens and others 1999 for U.S. numbers). The elasticities become even more strongly negative if the pension is assumed to be split over more household members.15 One reason for the large magnitude is likely the very low employ- ment rates in the first place, which make the marginal return to search quite small, in effect lowering the cost of leisure. Effects on Men and Women These regressions are also estimated separately on prime-age African men and women (table 3). More pension income significantly reduces both working hours and employment rates among prime-age men. More pension income is also as- sociated with fewer working hours for women, although the effect is smaller (–0.01 versus –0.015). Moreover, there is no apparent adjustment of female labor supply on the extensive margin (panel B, column 4). The only labor supply vari- able that does not appear to be significantly affected by the presence of eligible elderly is again female employment status (columns 2 and 5). Although the point estimate is negative, it is not statistically significant. In the preferred specifica- tion (columns 3 and 6) the effect on hours worked is much larger for men (2.2) than for women (1.3). Calculations similar to these (and assuming that men and women earn similar incomes) yield an elasticity of –0.66 for men and –0.43 for 12. One implication is that household income net of the pension declines when pension income in- creases. This can be verified in the household-level data by studying the effect of pension income on total nonpension household income. The iv specification finds that nonpension income goes down by about 1.05 rand for each extra rand of pension money. Jensen (1998) explores another channel, the decline in remittance income, through which the social pension can affect nonpension income. 13. Equal sharing among all prime-age individuals does not occur in practice, as is shown later in the article. 14. The scaling is on hours conditional on working because the effect on the work or not-work decision will be considered separately. 15. How reasonable is the assumption that prime-age people receive the full pension income? Re- sults reported later show that women respond less to pension income, suggesting that men get a dispro- portionate share. Duflo (1999) shows that the social pension improved the anthropometric status of girls under age five, suggesting that some of the pension income is spent on children. 36 Table 3. Effect of Old-Age Pension Income on Working Hours and Employment Status of 16- to 50-Year-Old African Men and Women Hours worked Employment status ols ols Pension Pension Pension Pension uptake eligibility iv a uptake eligibility iv Variableb (1) (2) (3) (4) (5) (6) Men Pension income × 1000 –15.13 (1.72) — –22.48 (2.72) –0.098 (0.034) — –0.201 (0.056) Household eligibility — –8.703 (0.849) — — –0.086 (0.018) — dummy 36 R2 0.163 0.166 — 0.234 0.234 — Women Pension income × 1000 –10.29 (1.11) — –13.27 (1.73) –0.018 (0.029) — –0.023 (0.043) Household eligibility — –4.810 (0.646) — — –0.013 (0.014) — the world bank economic review, vol. 17, no. 1 dummy R2 0.107 0.100 — 0.178 0.178 — Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size is 2,532 for men and 3,794 for women. Other covariates included in the regressions are a quartic in age, a dummy variable for having completed at least eighth grade, 14 province indicators, 3 location indicators (urban, rural, and metro), household size, number of household members ages 0–5, 6–15, 16–18, 19–21, and 22–24. aPension income is instrumented with the number of age-eligible women in the household and the number of age-eligible men in the household. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. Bertrand, Mullainathan, and Miller 37 women on the hours dimension and elasticities of –0.98 for men and –0.14 for women on the employment dimension. In regressions not reported here (but available from the authors) the drop in employment for men is explored more closely. The regressions examine whether the “missing� working men have entered a phase of unemployment or whether they have dropped out of the labor force. No difference in unemployment prob- abilities is found for eligible and noneligible households. Rather, the missing working men appear to have left the labor force. Moreover, there is no sign that the social pension increases the probability of discouragement. IV. Possible Confounding Effects The study next looks at the possibility that the estimates of the pension effects might be biased in that they attribute to the pension the effects of other, unob- served differences, or that they capture some other behavioral changes induced by the program rather than a decrease in labor supply. Direct Effect of the Presence of Elderly in a Household A primary concern about the results reported here is that individuals living in pension-eligible households might be systematically different from individuals liv- ing in pension-ineligible households. For example, the prime-age men and women living in eligible households are slightly younger than their counterparts in noneligible households. Furthermore, pension-eligible households are larger on average. It is conceivable that prime-age men living with older individuals are less qualified for work, less willing to look for work, or in some other way less likely to find work. If this is the case, then the estimates of the pension’s effect are biased because they attribute the effect of these unobserved differences to the pension. Several approaches are used to address this possibility. First, the nonlinearity in pension receipt as a function of the elder household member’s age is exploited to better separate the pension’s effect from these confounding factors. The pen- sion program rules predict a specific form for these nonlinearities: the presence of a woman older than 60 or of a man older than 65 should have large effects. There are no obvious reasons to expect such specific nonlinearities at these two age thresholds if the estimates are capturing a general impact of the presence of elderly people on the labor supply of younger household members. To examine how working hours for prime-age individuals living in a three- generation household are affected by the presence of elderly in different age groups, the impact on prime-age labor supply of living with eligible elderly is first compared with the impact of living with noneligible elderly (a dummy vari- able is used for the presence in the household of a woman between age 50 and 60 or a man between age 50 and 65). The presence of a noneligible elderly per- son in the household has neither a statistically nor an economically significant impact on prime-age working hours (table 4). But as already demonstrated, liv- ing with an eligible elderly person has a dramatic effect on working hours. 38 Table 4. Effect of the Presence of Elderly on Hours Worked by 16- to 50-Year-Old Africans Variable (1) (2)a (3) (4) (5) (6) (7) Eligible elderly in household –6.79 (0.64) — — — — — — Noneligible elderly –0.46 (0.63) — — — — — — in household Persons in household 50–55 — –0.42 (0.50) –0.22 (0.50) –0.40 (0.50) –0.21 (0.50) — — (n5055) Women in household 50–55 — — — — — –0.65 (0.67) –0.49 (0.67) (n5055f) Men in household 50–55 — — — — — –0.23 (0.96) –0.03 (0.95) (n5055m) Persons in household 55–60 — –0.23 (0.70) –0.09 (0.71) –0.22 (0.70) –0.08 (0.71) — — 38 (n5560) Women in household 55–60 — — — — — –0.19 (1.04) –0.04 (1.04) (n5560f) Men in household 55–60 — — — — — –0.62 (0.86) –0.50 (0.87) the world bank economic review, vol. 17, no. 1 (n5560m) Persons in household 60–65 — –2.54 (0.58) –2.41 (0.58) –2.53 (0.57) –2.41 (0.57) — — (n6065) Women in household 60–65 — — — — — –2.95 (0.94) –2.81 (0.93) (n6065f) Men in household 60–65 — — — — — –1.16 (1.41) –1.03 (1.41) (n6065m) n6065m × deviation from –7.47 (3.86) –7.43 (3.88) eligibility rule in regionb Persons in household over 65 — –5.37 (0.51) –5.21 (0.53) — — — — (n65p) Persons in household 65–70 — — — –5.17 (0.56) –5.03 (0.72) — — (n6570) Women in household 65–70 — — — — — –6.67 (0.86) –6.52 (0.88) (n6570f) Men in household 65–70 — — — — — –3.85 (1.19) –3.74 (1.20) (n6570m) Persons in household over 70 — — — –5.49 (0.56) –5.31 (0.57) — — (n70p) Women in household over 70 — — — — — –7.47 (0.72) –7.28 (0.71) (n70pf) Men in household over 70 — — — — — –2.87 (1.13) –2.76 (1.14) (n70pm) Household members over 50 — — –1.05 (0.69) — –1.04 (0.69) — –0.94 (0.68) with health problems R2 0.119 0.124 0.125 0.124 0.125 0.129 0.130 Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size in all regressions is 6,326. Other covariates included in regression are a quartic in age, a dummy for sex, a dummy for completion of at least 8th grade, 14 province indicators, 3 location indicators (urban, rural and metro), household size, number of household members ages 0–5, 6–15, 16–18, 19–21, and 22–24. 39 aTests of equality of coefficient below and above eligibility threshold: Column 2: n5055 = n6065 (p = 0.004), n5055 = n65p (p = 0.000), n5560 = n6065 (p = 0.009), n5560 = n65p (p = 0.000). Column 3: n5055 = n6065 (p = 0.003), n5055 = n65p (p = 0.000), n5560 = n6065 (p = 0.009); n5560 = n65p (p = 0.000). Column 4: n5055 = n6065 (p = 0.004), n5055 = n6570 (p = 0.000), n5055 = n70p (p = 0.000), n5560 = n6065 (p = 0.009), n5560 = n6570 (p = 0.000), n5560 = n70p (p = 0.000). Column 5: n5055 = n6065 (p = 0.003), n5055 = n6570 (p = 0.000), n5055 = n70p (p = 0.000), n5560 = n6065 (p = 0.009), n5560 = n6570 (p = 0.000), n5560 = n70p, (p = 0.000). Column 6: n5055f = n6065f (p = 0.020), n5055f = n6570f (p = 0.000), n5055f = n70pf, (p = 0.000), n5055m = n6570m (p = 0.011), n5055m = n70pm (p = 0.045), n5560f = n6065f, (p = 0.030), n5560f = n6570f (p = 0.000), n5560f = n70pf (p = 0.000), n5560m = n6570m (0.023), n5560m = n70pm (p = 0.100), n6065m = n6570m (p = 0.131), n6065m = n70pm (p = 0.379), n5055f = n6065f (p = 0.018), n5055f = n6570f (p = 0.000), n5055f = n70pf. Column 7: (p = 0.000), n5055m = n6570m (p = 0.009), n5055m = n70pm (p = 0.037), n5560f = n6065f, (p = 0.029), n5560f = n6570f (p = 0.000), n5560f = n70pf (p = 0.000), n5560m = n6570m (p = 0.022), n5560m = n70pm (p = 0.098), n6065m = n6570m (p = 0.129), n6065m = n70pm (p = 0.378). bDeviation from eligibility rule in region, the fraction of households with men 60–65 years old and no eligible elderly who are receiving a social pension in the region. This variable ranges from 0 to 0.67. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. Bertrand, Mullainathan, and Miller 39 40 the world bank economic review, vol. 17, no. 1 To further refine this finding, regressors are added for the number of people in each of the following age categories: 50–55, 55–60, 60–65, and 65 and older (column 2). The coefficients clearly show a negative effect of the presence of elderly between 60 and 65 years old and an even stronger negative effect of the presence of elderly older than 65. On the other hand, the presence of elderly people between ages 50 and 55 and between ages 55 and 60 seems to have nei- ther an economically nor a statistically significant impact on working hours among prime-age individuals. Moreover, the test statistics clearly reject the hy- pothesis that the pre-eligibility coefficients are equal to the posteligibility coeffi- cients (see table 4). Although these results provide some compelling evidence, it is still possible that the effect of the elderly person’s age has an independent, nonlinear effect. Most notably, the very old are more likely to have health problems and to re- quire some assistance at home. This may cause prime-age individuals who live with them to reduce their labor force participation to provide home care.16 The second strategy tries to deal directly with this problem. The survey asks respondents to list any household member who has been sick or injured over the past two weeks, “including people who have some form of permanent injury, disability, or ailment.� To investigate whether such health problems display the same nonlinearity as the pension rule, the probability of having some health problem was regressed on 10 dummy variables for age and gender groups of elderly for the entire data set (the results are not detailed here).17 Although people 50–55 years old appear healthier than people older than 55 (this is true for both men and women), there is no statistically significant difference in the probabil- ity of being sick for people age 55–60 and people older than 60. Hence, if there is any age discontinuity in health status, it appears to occur before the age of pension eligibility. The number of elderly household members who report health problems are then included in the employment regression (column 3 of table 4). Though the coefficient on health problems is negative (each sick elderly person is associated with an hour less of work) and marginally significant, it does not affect the pen- sion coefficients. These results exhibit the same discontinuity pattern as they do in column 2, suggesting that the health status of the elderly does not drive the results. Columns 4 and 5 replicate the specifications in columns 2 and 3 but further break down the number of people older than 65 into number of people age 65– 70 and number of people older than 70. The results are unchanged. The coeffi- 16. At first glance, this story seems inconsistent with the fact that men reduce their work more than women do. If women provide the main input in home care, they ought to reduce their work hours more. One could, however, claim that women are expected to both care for the elderly and work, whereas men will either work or take care of the elderly. 17. The dummy variables are women age 50–55, women age 55–60, women age 60–65, women older than 70, and the equivalent age groups for men. Bertrand, Mullainathan, and Miller 41 cients on all the age categories below the eligibility threshold are not statistically significant different from zero. The coefficients on all the age categories above the eligibility threshold are significant and negative. Moreover, the test statistics show that the hypothesis of equality of the coefficients below and above the eli- gibility threshold can be rejected (see table 4). The third strategy exploits regional differences in how the pension program is implemented. In certain areas authorities deviated from the rule that men are eligible at a later age than women, informally extending the pension to men between 60 and 65 years old.18 If the results are in fact due to the pension, then in the regions that deviated from the official rule the presence of men 60–65 would be expected to affect household labor supply. Because of the informal nature of the extension, administrative data are not available on which areas altered the rule, but a proxy can be computed from the data for the fraction of households with men 60–65 years old and no other age-eligible elderly who report receiving some pension income.19 This fraction ranges from 0 in the most compliant prov- ince to 0.67 in the least compliant province. This fraction is interacted with a dummy variable for the number of men 60– 65 years old in the household (column 6). All the age groups of column 4 are further disaggregated by gender categories. The results are striking. None of the pre-eligibility coefficients are statistically different from zero. All the posteligibility coefficients are significant and negative. The direct effect of number of men 60– 65 years old (the effect in the provinces that do not deviate from the eligibility rule) is not statistically different from zero. The interaction term between devia- tion from the eligibility rule and number of men 60–65 years old is negative and significant. Finally, 10 of the 12 test statistics reject the assumption of equality between pre-eligibility coefficient and posteligibility coefficient. The same results hold after controlling for the number of elderly with health problems (column 7). These results suggest that the extension to pre-eligible men does in fact cor- relate with the pension’s estimated effect, bolstering the argument that the re- sults are not capturing spurious effects of age. In summary, the results in this section provide multiple pieces of evidence to suggest that what is being identified is a causal effect of the pension and not an independent effect associated with living with elderly people.20 18. As Case and Deaton (1998) report, the age differential in pension eligibility is technically un- constitutional and under revision at the central government level. Certain local authorities might have already gone ahead with age equalization by 1993. 19. Remember that pension income is observed not at the individual level but at the household level. 20. In a final attempt to account for possible confounding factors associated with the age of the elderly, the relationship between employment and pension eligibility prior to the bulk of the reform of the social pension program in South Africa was examined using the 1991 Population Census. This cross- sectional household survey was conducted prior to the major extension of the social pension to African households. Although the process of racial equalization of the social pension has been under way since the early 1990s, only after 1992 were the means tests unified, racial parity in benefits levels achieved, and new technologies introduced to improve benefit delivery. Thus, although the 1991 census was not administered prior to the beginning of reform, it was administered at a time when the pension was far 42 the world bank economic review, vol. 17, no. 1 Is the Labor Supply Response Real? Even taking as given that the regressions are identifying some causal effect of the pension, there might still be concern that the results are capturing other be- havioral changes induced by the program, rather than a decrease in labor supply. Bertrand and others (2000) present a thorough investigation of such alternative interpretations, which are simply summarized here. First, they find no evidence that what is being observed is a shift to casual or farm employment, which might be more difficult to measure than regular forms of employment. Reported casual working hours actually decline, and the level of self-employment does not change. The level of home production activities, such as agricultural crop production or livestock production, does not change either. There is also no evidence for the related possibility that prime-age individuals living with pensioners are investing more in human capital. In fact, it is the older rela- tives of pensioners, not the school-age ones, who show the largest drop in work- ing hours. Another alternative is that the results are simply picking up on migration behavior. The pension may make the unemployed more likely to move in with the pensioners or the employed more likely to move out. There is no indication that migration patterns and family size are significantly affected by these variables. V. Distribution of Effects The results so far provide some evidence of a redistribution of the pension to- ward prime-age workers in the household. This section pushes the analysis a step further and asks whether the South African experiment can teach us more about how resources are allocated and collective labor supply decisions made within these extended families. Test of Income Pooling There are several prominent theoretical models of resource allocation within households. One, known as the common preference model, assumes that house- holds are best described as maximizing a single utility function (Samuelson 1956).21 A central result from the common preference model is that money is money. Which member of the household gets the marginal dollar of nonlabor income will affect neither the ultimate consumption level nor the leisure choice less generous and accessible to Africans. No evidence was found that the large negative employment effects in pension-eligible households (–6.8 percent) using the 1993 survey data are present in the 1991 data. Though there is some negative effect associated with the presence of an age-eligible person in 1991 (not surprising, given that a limited pension program was already in place), the effect is less than a quarter that of the 1993 pension. These census results are discussed in more detail in Bertrand and others (2000). 21. The common preference approach can be motivated either through the assumption of a family consensus, as in Samuelson (1956), or through the assumption of altruistic behavior, as in Becker’s “rotten kid� theorem (Becker 1974, 1981). Bertrand, Mullainathan, and Miller 43 of each household member. This result holds even in the presence of differential altruism across individuals. The individuals who get more resources receive the greatest weight in the joint household utility function. Another important set of models rejects the idea that families can be reduced to a single optimizing agent. These models assume that household members have distinct preferences, and the models look at how bargaining between members affects the allocation of resources. Most often the bargaining consists of a Pareto efficient process, such as a Nash bargaining model between the different par- ties.22 A central feature of these bargaining models is that the strong fungibility result found in the common preference model no longer holds: who gets the money matters. The income controlled by each household member influences the bargained outcome. Moreover, the higher the bargaining power of a house- hold member, the greater the resources that member will receive. The social pension program in South Africa provides an unusual opportunity for an experiment that separates common preference and bargaining models of the family. As mentioned earlier, the social pension, although in theory means- tested, is in practice mainly a lump-sum transfer for African households. Hence, the pension transfer, especially when instrumented for by the presence of age- eligible elderly, does not depend on earned income or other possible choice vari- ables in the household resource allocation decision. One can therefore test for income pooling by asking whether pension transfers made to elderly women have the same effect on prime-age labor supply as pension transfers made to elderly men, holding family composition constant. The findings reported in table 4 already suggest that an elderly woman’s pen- sion income may have a larger negative effect on prime-age labor supply than an elderly man’s pension income. The coefficients on number of women above the eligibility threshold are systematically larger (in absolute value) than the coefficients on number of men above the eligibility threshold. When the work- ing hours of both men and women are regressed on the standard set of geographic, individual, and family controls, and regressors are added for number of age- eligible women and age-eligible men, the coefficient on number of eligible women is more than twice that on number of eligible men (column 1 of table 5). Such differences are not due to any measurement error in the number of eli- gible men. Even when accounting for the fact that some men 60–65 years old also receive pension income in certain South African provinces (column 2), the coefficient on number of men older than 65 is still only half that on number of women older than 60.23 Finally, this finding still holds after controlling for the 22. Several researchers, such as Manser and Brown (1980), McElroy and Horney (1981), and Lundberg and Pollak (1993), have developed cooperative Nash bargaining models of intrahousehold resource allocation. Chiappori (1992) produces a far more general model that includes all Pareto effi- cient bargaining models. 23. The coefficient on the interaction term between number of men 60–65 and deviation from the eligibility rule is insignificant though negative. The rise in standard error is due to the fact that each of the age categories is not included separately. 44 the world bank economic review, vol. 17, no. 1 Table 5. Old-Age Pension and Pooling of Resources Prime-age living with exactly All prime-age 1 elderly woman in 3-generation households and 1 elderly man Variablea (1) (2) (3) (4) Number of women over 60 –5.02 (0.58) –5.13 (0.58) –5.13 (0.57) –3.89 (1.44) Number of men over 65 –2.32 (0.87) –2.55 (0.87) –2.54 (0.88) –0.71 (1.47) Number of men 60–65 — –1.13 (1.46) –1.12 (1.43) — Number of men 60–65 × — –5.31 (4.06) –5.10 (4.05) — deviation from elibility ruleb Number of women 16–50 — — 0.089 (0.19) — Number of men 16–50 — — –0.14 (0.22) — Number of women 0–16 — — –0.41 (0.17) — Number of men 0–16 — — 0.01 (0.18) — R2 0.118 0.120 0.122 0.120 Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size is 6,326 in columns 1–3 and 1,471 in column 4. aOther covariates included in all regressions are 14 province indicators and 3 location indicators (urban, rural, and metro). Also included in columns are a quartic in age, a dummy for sex, a dummy for having completed at least eighth grade, household size, number of household members ages 0–5, 6–15, 16–18, 19–21, and 22–24. bFraction of households with men 50–65 years old and no eligible elderly who are receiving a social pension in the region. This variable ranges from 0 to.67. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. number of prime-age men, number of prime-age women, number of male chil- dren and number of female children (column 3). In other words, the fact that women’s pension money reduces labor supply more than men’s pension money cannot be explained by any systematic difference in the number and gender com- position of the nonelderly in households with eligible women and in households with eligible men. This first finding appears inconsistent with pooling of resources within the household and builds some preliminary support against a common-preference model of collective labor supply. The marginal dollar of pension income received by an elderly woman reduces labor supply more than the marginal dollar of pension income received by an elderly man. However, this first finding is not conclusive. It does not account for the possibility that the marginal rand of pen- sion income going to an elderly woman may have to be distributed among a different set of household members then the marginal dollar of pension money going to an elderly man. Of primary concern here is the possibility that old women might have a lower weighting than old men in a household utility function. If that is the case, and assuming that households with eligible women have more elderly women than households with eligible men (a very likely event), the finding could still be recon- ciled with the common preference model. To deal with this concern the sample is Bertrand, Mullainathan, and Miller 45 restricted to the set of households that have exactly one elderly woman (older than 50) and one elderly man (also older than 50).24 In this subset of households the marginal rand of pension income, whether from a female or a male pensioner, will be reallocated among a fixed number of elderly of each gender. Replicating the specification of column 1 on that subset of households still yields a much stronger negative labor supply response for pension money going to women than pension money going to men (column 4 of table 5). The marginal rand of pension income going to a female pensioner reduces labor supply by about three times as much as the marginal rand of pension income going to a male pensioner. The coefficient on number of age-eligible men is, however, less precisely estimated. Although the number and gender composition of the elderly are forced to be the same in this restricted sample, it could be that the number and gender com- position of the nonelderly systematically varies with the gender of the pensioner. For the subset of households with exactly one woman older than 50 or one man older than 50, no statistically significant difference was found in the number and gender of prime-age individuals between households with a female pensioner and households with a male pensioner. However, households with a female pensioner had slightly more children than households with a male pensioner. This last dif- ference suggests that the marginal rand of pension money needs to be split among slightly more individuals when the pensioner is a woman. Hence, under the com- mon preference model, this could only lead the coefficient on eligible women to be smaller in absolute value than the coefficient on eligible men if children re- ceive any weighting in the family utility function. This is exactly the opposite of the results found. In summary, the results in table 5 strongly point toward rejecting the income- pooling hypothesis for African extended families. A marginal rand of pension income has a drastically different effect on prime-age labor supply depending on whether the pension-earning person is a man or a woman. This still holds true when households differ only in the gender of their pensioners and not in the number, age, and gender composition of their members. Who Benefits from the Old-Age Pension? Testing for income pooling is only one way to look at how money is distributed inside the household. Another would be to examine who are the biggest benefi- ciaries of the reallocation of resources. Is the pension money evenly distributed among all the prime-age individuals in the extended family or are certain family members able to reduce their labor supply more than others? Answering this question involves estimating the standard regression of hours worked on the pension variable, but this time interacting the pension variable with several de- mographic characteristic (table 6).25 24. There are few households that contain two or more elderly persons of both genders. 25. For reasons of space results are reported in table 6, mainly for the iv specification. 46 Table 6. Distribution of the Effect of Old-Age Pension Income on Working Hours of 16- to 50-Year-Old Africans Variablea (1) (2)b (3) (4) (5) (6) (7) Pension income × 1000 –21.04 (2.51) — –14.09 (1.85) –15.55 (1.63) 25.48 (3.45) 39.35 (12.37) 20.79 (3.66) Pension income × 1000 × female 9.05 (2.21) — — — — — 7.10 (2.47) Women over 60 — –6.98 (0.81) — — — — — Men over 65 — –2.73 (1.03) — — — — — Women over 60 × female — 3.23 (0.75) — — — — — Men over 65 × female — 0.71 (0.89) — — — — — Pension income × 1000 × — — –7.42 (3.02) — — — — 4th grade or less Pension income × 1000 × matric — — — –0.07 (3.93) — — — or more 46 Pension income × 1000 × age — — — — –1.53 (0.14) –2.54 (0.91) –1.49 (0.14) Pension income × 1000 × age2 — — — — — 0.02 (0.02) — Pension income × 1000 × oldest — — — — — — –9.08 (5.00) prime-age man in household the world bank economic review, vol. 17, no. 1 Note: Numbers in parentheses are standard errors. Standard errors are corrected to allow for group effects within survey household clusters. Sample size is 6,326 in columns 1–6 and 6,189 in column 7. aOther covariates included in regression are a quartic in age, a dummy variable for gender, a dummy variable for having completed at least 8th grade, 14 province indicators, 3 location indicators (urban, rural, and metro), household size, number of household membersages 0–5, 6–15, 16–18, 19–21, and 22– 24. Columns 3, 4, and 7 also include a dummy variable for “4th grade or less,� a dummy variable for “matric or more,� and a dummy variable for “oldest prime-age man in the household.� bAll columns except column 2 represent iv results. In the iv specifications, pension income and the interactions of pension income with the other variables of interest are instrumented. Source: All variables are from the World Bank/South African Development Research Unit survey, August–December 1993. Bertrand, Mullainathan, and Miller 47 As the results in table 3 have already suggested, the pension reduces the labor supply of prime-age men more than that of prime-age women. The effect of the social pension on women’s labor supply is about half the effect on men’s (col- umn 1 of table 6), implying that prime-age women benefit less from the social pension than do prime-age men. To see whether this effect depends on the gen- der of the pensioner, the female dummy variable is interacted with the number of eligible women and the number of eligible men in the household (column 2). The presence of an additional male pensioner in a household does not have a statistically different effect on male and female labor supply, whereas the pres- ence of an additional female pensioner benefits prime-age men more than it does prime-age women. That means that the greater effect of the pension on prime- age men occurs only when the pensioner is a woman. Education might also influence who benefits more from the pension transfer. On one hand, individuals with higher educational attainment presumably have more outside options, which could increase their threat points when bargaining over resources with other family members. On the other hand, at a given level of redistribution, individuals with the lowest market wages may give up their job first. If educational attainment is positively correlated with market wage, the least educated workers would be expected to reduce their labor supply most. The differential effect of the pension by education group thus appears to be an em- pirical question. The prime-age men and women in the sample who have not completed fourth grade (about a quarter in each group) reduce their labor sup- ply by about 50 percent more than the individuals who have completed at least the fourth grade (column 3). There is no difference in labor supply response, however, between individuals who have completed at least the matric (10th grade) and those who have not. Thus, it appears that the labor supply response is stron- gest among the least skilled, probably because they face such unattractive labor market options to start with. Age could also affect the size of the labor supply response of prime-age house- hold members. The social pension depresses labor supply more as prime-age men get older (column 5). Allowing for a quadratic relation between pension money and age enables an assessment of whether the effect peaks at any point over the range of working ages. The effect of age appears mostly linear and does not peak before 50 years of age (column 6).26 Another question is whether it is absolute or relative age that affects intra- household redistribution. More precisely, does the special position that oldest sons are believed to hold inside the family result in the oldest prime-age man in the household receiving more pension money than other household mem- bers? After the differential effects of the pension by age and gender are con- trolled for, the results show that the oldest man in the household reduces his 26. These results also undercut the argument that household members reduce their labor supply to get an education. 48 the world bank economic review, vol. 17, no. 1 labor supply more than other household members (column 7).27 These results support the view that oldest sons receive more resources in extended families. After the direct effect of own age and gender on resource distribution are ac- counted for, the results show that the oldest man reduces his labor supply by about 50 percent more than other men in the household and about 70 percent more than women. The results in this section on the distribution of effects can be understood in light of the bargaining models of household resource allocations. First, the ob- served differences in redistribution can be attributed to differences in bargain- ing power. Men’s labor supply responds more to pension income because men have more power inside the household. That the male–female differential is largest when the pensioner is a women is suggestive of a situation in which dominant males capture resources. When the pensioner is a man, the ability of prime-age men to capture resources is diminished and so is the male–female differential in labor supply response. The age results are in line with this picture. The oldest male seems most capable of capturing household resources. Alternatively, differences in altruism could explain the patterns observed in table 6. Perhaps pensioners care more about men. To fit the results here, female pensioners would have to care the most about prime-age male household mem- bers. Moreover, pensioners’ altruism would have to be strongest toward their oldest prime-age male children. Even if this pattern of altruism does not seem particularly intuitive, it nevertheless provides another lens for interpreting the findings. VI. Conclusion With improving health conditions and lengthened life expectancy in many de- veloping economies, governments will soon have to introduce full-fledged social programs to provide for the needs of a growing elderly population. Before simply replicating the type of programs that industrial countries have installed, policy- makers in developing economies would do well to consider how different living arrangements could interfere with their social objectives. Though the elderly in industrial countries may often live on their own, multigeneration households prevail in developing countries. The South African pension program provides a way to understand the effects of such targeted programs when extended family links are strong. The South African government’s pension program was introduced as a way to improve living conditions for older individuals who are no longer in the labor force and who do not have access to a private pension. The vast majority of the older Africans in South Africa participate in this program. This article provides some evidence that, in practice, at least part of the cash transfers targeted for the 27. The sample size is slightly smaller here than in the basic regression because the 2 percent of households that have only one prime-age individual are excluded from the sample. Bertrand, Mullainathan, and Miller 49 elderly ends up in the hands of a group that was not originally targeted: prime- age men and women who live with the pensioners. The results reported here indicate that African household members 16–50 years old reduce their labor supply when they live with pension beneficiaries. Hence, because of intra- household redistribution, a program designed for a group that is out of the labor force unexpectedly altered the labor supply of a nontargeted group. Moreover, the study relates this labor supply response to standard theories of intrahousehold resource allocation and collective labor supply choice. The dif- ferent labor supply impact of money from male and female pensioners suggests that a common preference model of family labor supply cannot adequately de- scribe the results and that some amount of bargaining takes place within these households. In general, older prime-age men, in particular the oldest prime-age man in a household, appear to be the biggest beneficiaries of the pension. Within the set of bargaining models of intrahousehold allocation of resources, this find- ing could be interpreted as evidence that these men have relatively more bar- gaining power or are being cared for more by other family members. References Altonji, Joseph, Fumio Hayashi, and Laurence Kotlikoff. 1992. “Is the Extended Family Altruistically Linked? Direct Tests Using Micro Data.� American Economic Review 82(5):1177–98. Ardington, Libby, and Frances Lund. 1994. “Pensions and Development: The Social Security System as a Complementary Track to Programs of Reconstruction and De- velopment.� University of Natal Centre for Social and Development Studies, Durban. Becker, Gary. 1974. “A Theory of Social Interaction.� Journal of Political Economy 82(6):1063–94. ———. 1981. A Treatise on the Family. Cambridge, Mass.: Harvard University Press. Bertrand, Marianne, Sendhil Mullainathan, and Douglas Miller. 2000. “Public Policy and Extended Families: Evidence from South Africa.� Working Paper, University of Chicago Graduate School of Business, Chicago. Browning, Martin, Francois Bourguignon, Pierre-André Chicappori, and Valérie Lechene. 1994. “Income and Outcomes: A Structural Model of Intrahousehold Allocation.� Journal of Political Economy 102(6):1067–96. Caelers, Di. 1998. “Noah Saves Khayelitsha’s Old Folk.� Independent, June 16. Case, Anne, and Angus Deaton. 1998. “Large Cash Transfers to the Elderly in South Africa.� Economic Journal 108(450):1330–61. Chiappori, Pierre-Andre. 1992. “Collective Labor Supply and Welfare.� Journal of Po- litical Economy 100(3):437–67. Duflo, Esther. 1999. “Child Health and Household Resources in South Africa: Evidence from the Old Age Pension Program.� Working Paper, Massachussetts Institute of Technology, Cambridge, Mass. Imbens, Guido, Donald Rubin, and Bruce Sacerdote. 1999. “Estimating the Effects of Unearned Income on Labor Supply, Earnings, Savings, and Consumption: Evidence from a Survey of Lottery Players.� nber Working Paper 7001. National Bureau of Economic Research, Cambridge, Mass. 50 the world bank economic review, vol. 17, no. 1 Jensen, Robert. 1998. “Public Transfers, Private Transfers and the ‘Crowding Out’ Hypothesis: Evidence from South Africa.� Working Paper, Harvard University, John F. Kennedy School of Government, Cambridge, Mass. Lund, Frances. 1992. “State Social Benefits in South Africa.� International Social Security Review 46(1):5–25. Lundberg, Shelley, and Robert Pollak. 1993. “Separate Spheres Bargaining and the Marriage Market.� Journal of Political Economy 101(6):988–1010. Lundberg, Shelley, and Robert Pollak. 1996. “Bargaining and Distribution in Marriage.� Journal of Economic Perspectives 10(4):139–58. Manser, Marilyn, and Murray Brown. 1980. “Marriage and Household Decision Making: A Bargaining Analysis.� International Economic Review 21(1):31–44. McElroy, Marjorie, and Mary Jean Horney. 1981. “Nash-Bargained Household Deci- sions: Towards a Generalization of the Theory of Demand.� International Economic Review 22(2):333–49. Ngoro, Blackman. 1998. “Retiring to Life in Shackland.� Independent, July 20. Samuelson, Paul. 1956. “Social Indifference Curves.� Quarterly Journal of Economics 70(1):1–22. Subramanian, Shankar, and Angus Deaton. 1991. “Gender Effects in Indian Consump- tion Patterns.� Sarvekshana 14(4):1–12. Van der Berg, Servaes. 1994. “Issues in South African Social Security.� World Bank, Washington, D.C.