77318 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2: 347-«7 Cnild Labors Child Schooling, and Their Interaction with Adult Labor: Empirical Evidence for Peru and Pakistan Ranj an Ray Using data from Peru and Pakistan, this article tests two hypotheses: there is a positive association between hours of child labor and poverty, and there is a negative association between child schooling and poverty. Both of these hypotheses are confirmed by the Pakistani data, but not by the Peruvian data. The reduction in poverty rates due to income from children's labor is greater in Pakistan than in Peru. The nature of interac- tion between adult and child labor markets varies with the gender of the child and the adult. In Peru rising men's wages significantly reduce the labor hours of girls, whereas in Pakistan there is a strong complementarity between women's and girls' labor markets. Both data sets agree on the positive role that increasing adult education can play in improving child welfare. In recent years academics, public officials, and the media have shown a growing interest in child labor. Although the International Labour Organisation's (ILO) estimates of labor force participation rates for children ages 10-14 years show a declining trend (ILO 1996a), in absolute terms the size of the child labor force is and will continue to be large enough to be of serious concern. According to the International Labour Organisation, in 1990 almost 79 million children between 5 and 14 years of age around the world were working full time (see Ashagrie 1993:16). Ashagrie (1993) was the first person to compile an international data set on child labor. His figure on the size of the child labor force (79 million) has since been revised upward to 120 million children (ILO 1996b and Ashagrie 1998). Including part-time workers as well, this estimate rises to 250 million. Although the estimate of the child labor force varies depending on how we define work, how we define a child, and how we collect data, few would disagree that child labor is a problem of gigantic proportions. The universal perception of child labor as a problem stems from the widespread belief that employment is destructive to children's intellectual and physical development, especially that of Ranjan Ray is with tbe School of Economics at the University of Tasmania. His e-mail address is ranjan.rayQutas.ediuiu. Part of this research was carried out during the author's study leave at Cornell University. The research was sponsored by the World Bank project "Intrahousehold Dedsionmaking, Literacy, and Child Labor.* The author is grateful to Kaushik Basu for comments and suggestions and to Geoffrey Lancaster for skillful and painstaking research assistance. He also acknowledges helpful comments from the editor and the anonymous referees on an earlier version of this article. © 2000 The International Bank for Reconstruction and Development /THE WORLD BANK 347 348 THE WORLD BANX ECONOMIC REVIEW, VOL. 14, NO. 2 young children. The danger is particularly serious for children who work in haz- ardous industries. Working prevents children from benefiting fully from school and may thereby condemn them to perpetual poverty and low-wage employ- ment. Moreover, many believe that child labor contributes to adult unemploy- ment in developing countries. According to an U.O study, "Child labour is a cause of and may even contribute to adult unemployment and low wages" (Bequele and Boyden 1988: 90). The trade lobby in industrial countries further argues that child labor gives developing countries an unfair cost advantage in international trade and that this constitutes a form of protection requiring corrective interna- tional action. For example, in a study of child labor in India's carpet industry, Levison and others (1996: abstract) find "a competitive cost advantage to hiring child labour with its magnitude relatively small for industrialised country sellers and consumers but relatively large for poor loom enterprise owners." Notwithstanding almost universal agreement that child labor is undesirable, there is wide disagreement on how to tackle the problem. This stems partly from the lack of awareness of—let alone consensus on—the causes of child labor and the consequences of banning it through legislation. Although a few studies, for example, Knight (1980) and Horn (1995), discuss mainly the qualitative features of child labor, most focus on its quantitative aspects, taking advantage of the increasing availability of good-quality micro-data.1 Within the empirical literature on child labor there has been a shift from mere quantification to econometric analysis of its determinants. This has. coincided with a widespread realization that simply banning child labor is unlikely to eradi- cate the problem and may even make it worse. As Knight (1980:17) notes, "When child labor is prohibited by law, the law cannot protect child workers since they legally do not exist." The view that we need to understand the key determinants of child labor in order to formulate effective policies against it underlies the re- cent econometric work. Early studies centered mostly on Latin America, but re- cent studies have extended the focus to countries in Africa and Asia.2 In this article I use data from Peru and Pakistan to examine the key determi- nants of child labor hours and the share of child and adult earnings in the household's total earnings. I focus on the similarities and dissimilarities between the Peruvian and Pakistani results. This study extends Ray (1998), which ana- lyzes child labor force participation, rather than child labor hours, and restricts attention to child1 employment, neglecting adult labor. Moreover, this article over- comes a significant limitation of Ray (1998) by estimating the equations for child labor hours separately for boys and for girls. 1. See Grootaert and Kanbur (1995) and Bare (1999) for survey* of the recent literature. 2. Examples include Psacharopoulos (1997) on Venezuela; Cartwright and Patrinos (1998) on Bolivia; Grootaert (1998) on Cdte dlvoire; Tienda (1979), Boyden (1988,1991), and Patrinos and Psacharopoulos (1997) on Peru; Salazar (1988) on Colombia; Patrinos and Psacharopoulos (1995) on Paraguay; Myen (1989) on South America; Bonnet (1993) on Africa; Jensen and Nielsen (1997) on Zambia; Addison and others (1997) on Pakistan and Ghana; Chaudhuri (1997) on India; Diamond and Fayed (1998) on Egypt; and Ray (1998) on Pakistan and Peru. See also the volumes edited by Bequele and Boyden (1988), Myen (1991), and Grootaert and Patrinos (1998). Ray 349 A key empirical result of this study is that the interaction between men's and children's laborj markets is different from the interaction between women's and children's labor markets. The importance of this distinction is sharpened by the fact that the responsiveness of child labor hours to changes in adult wages is different in the Peruvian and Pakistani data sets. The case for distinguishing be- tween how men's and women's labor interact with child labor can be traced to Grant and Hamermesh (1981), who find that children and white women were substitutes, whereas children and white men were complements in production in the United States.3 Another significant motivation of this article is to test the "luxury axiom" introduced by Basu and Van (1998:416): "A family will send the children to the labor market only if the family's income from non-child-labor sources drops very low." In view of the key role that this axiom plays in Basu and Van's analysis of child labor, an empirical evaluation of it has some policy significance. The test here throws light on the important question of whether poverty is the key deter- minant of child labor, as it is widely believed to be. Using data from more than one country is a good way to test whether this axiom is universal. I chose the Peruvian and Pakistani data sets on the basis of two consider- ations. First, because the World Bank collected data in both these countries through its Living Standards Measurement Study (LSMS) program, the structure and cov- erage of these data sets are sufficiently similar to allow meaningful comparisons. Second, Peru and Pakistan make an interesting comparison because of their geo- graphic distance from each other as well as their cultural, economic, and demo- graphic diversity. For example, per capita gross national product (GNP) in Peru is more than twice that in Pakistan, and its population is only one-fifth the size.4 The two countries also differ in family size, gender composition of the house- hold, and the education and work experience of women and children. Cross- country comparisons, especially those involving different cultures and continents, enable us to better understand the different types of policies needed in specific regional contexts. I also examine the role of community infrastructure by including some com- munity variables in the regressions. The results are of considerable practical sig- nificance since the community variables may be useful instruments for improving children's welfare. I. PRINCIPAL FEATURES OF THE DATA The data on child labor come from the 1994 Peru living Standards Measure- ment Survey (PLSS) and the 1991 Pakistan Integrated Household Survey (PIHS). 3. See Diamond and Fayed (1998) for more recent evidence in favor of distinguishing between men's and women's labor. Note also that each of the case studies in the recent volume edited by Grootaert and Patrinos (1998) wragniT^f this distinction. I am grateful to the referee for drawing this volume to my attention. 4. See Lancaster, Ray, and Valenzuela (1999: table 3). 350 THE WORLD BANK ECONOMIC REVIEW, VOL 14, NO. 2 These surveys were conducted jointly by the respective governments and the World Bank as part of the LSMS surveys carried out in a number of developing coun- tries.5 The LSMS surveys were designed to provide policymakers and researchers with the individual-, household-, and community-level data needed to analyze the impact of policy initiatives on living standards. The PLSS covers 3,623 households, and the PIHS covers 4,800 households. The Peruvian sample contains information on child labor and child schooling for 5,231 children ages 6-17 years, and the Pakistani data set includes 5,867 obser- vations on children ages 10—17 years. Some of these observations could not be used, however, because of their poor quality. To construct the data set on child wages, I combined information on income and work from a number of sources. The data on wages and labor hours pertain to children who are involved in full-time labor outside the home, for which they receive direct or indirect cash payments. I calculate the wages of children who work on family farms from the information on farm income and from die corre- sponding adult and child labor hours contained in the surveys. The regression estimates are robust to the changes in children's wage rates implied by alternative a priori apportionments of farm income between adults and children. Since a central motivation of diis study is to investigate the economic determi- nants of child labor, such as poverty and, especially, child wages, I do not con- sider in die substantive part of diis article work that involves purely household chores nor forms of child labor that do not receive, explicidy or implicidy, cash remuneration. This limits die scope of die analysis somewhat, but makes it con- sistent with the lLO's definition of child laborers as "economically active" chil- dren (see Ashagrie 1993). The data set from Peru does not provide information on hours children spent on unpaid domestic work, and I dierefore use all of the available information on child labor. In Pakistan the survey designers assumed diat boys were not in- volved in domestic work, but household chores accounted, on average, for more than 90 percent of girls' total labor hours. I dius test the sensitivity of die regres- sion results when unpaid, domestic hours are included as child labor for die case of Pakistani girls. In bodi Peru and Pakistan rates of labor market participation for children increase widi age (tables 1 and 2). These rates are similar for both countries. In the case of schooling, however, enrollment peaks at around 13 years in Peru. In Pakistan enrollment peaks earlier, at 11 years, and falls continuously to alarm- ingly low levels, especially for older girls. The gender picture is similar between the two countries with respect to child labor, whereas the enrollment rates of Peruvian children in all age groups are consistendy higher than those of dieir Pakistani counterparts. Pakistan's lower enrollment rates reflect die lack of good schools, compared with Peru. As Basu (1999) points out, the provision of good-quality schooling 5. See Grosh and Glewwe (1995) for an overview and general description of the LSMS data sets. Ray 3S1 Table 1. Participation Rates of Peruvian Children in Employment and in Schooling , (percent) Employment Schooling Age Boys Girls Total Boys Girls Total 6 7.9 11.6 9.6 90.5 89.3 89.9 7 12.9 11.8 12.4 93.1 94.6 93.8 8 17.6 11.6 14.3 95.5 95.9 95.7 9 18.5 17.1 17.8 98.1 99.5 98.8 10 29.4 22.1 25.8 97.2 97.1 97.2 11 31.8 21.7 27.0 98.3 96.7 97.5 12 37.7 27.0 32.2 95.5 94.8 95.1 13 32.0 27.3 29.5 96.1 88.1 91.9 14 48.7 32.4 40.6 89.3 90.1 89.7 15 51.8 32.7 42.2 88.2 83.2 85.7 16 46.1 34.9 40.4 82.7 74.4 78.5 17 57.1 27.9 42.6 63.9 58.7 61.3 All ages 31.8 22.7 27.3 90.9 89.0 90.0 Source: 1994 Peru Living Standards Measurement Survey. Table 2. Participation Rates of Pakistani Children in Employment and in Schooling (percent) Employment Schooling Age Boys Girls Total Boys Girls Total 10 14.9 18.7 16.7 77.3 51.1 64.5 11 16.1 19.6 17.7 82.2 54.8 69.6 12 25.4 22.8 24.2 73.5 49.0 62.4 13 30.3 21.3 25.6 72.1 45.3 58.1 14 36.3 28.3 32.2 66.8 39.0 52.6 15 39.8 29.8 35.0 56.9 33.4 45.7 16 51.2 26.7 39.0 50.7 28.1 39.4 17 48.4 25.8 38.9 48.8 28.2 40.1 All ages 31.3 23.9 27.8 67.2 42^ 55.2 Source: 1991 Pakistan Integrated Household Survey. can play a big part in reducing child labor. Moreover, strict Islamic laws that keep women at home may explain the sharp fall in enrollment among older Paki- stani girls, especially those in the age group 11-16 years. In that age group the employment and school enrollment rates of Pakistani boys move sharply in op- posite directions, suggesting that, unlike Peruvian boys, Pakistani boys drop out of school completely to enter the labor market. This behavior is consistent with Weiner's (1996: 3007) observation on Indian child labor: "Most of the 90 mil- lion children not in school are working children." Note that some of Weiner's "working children" are not included here as child laborers because they do not receive wages.6 6. Cartwrigfat and Patrinos (1998), for example, classify child workers who do not attend school and are not formally employed as "home care" workers. 352 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 It is possible to compare other key characteristics of the two data sets by look- ing at sample means (table 3). A typical Pakistani household contains more chil- dren than a typical Peruvian household, although the rates of child poverty are similar in the two countries. With the poverty lines defined as 50 percent of the median nonchild household income per adult equivalent in the sample, the per- centage of children living in poverty is similar in the two countries: 30.4 and 29.4 percent for boys and girls in Peru, and 27.1 and 25.6 percent for boys and girls in Pakistan. If we include children's earnings in household income, these figures fall to 29.3 and 29.0 percent for boys and girls in Peru and to 23.4 and 23.7 percent for boys and girls in Pakistan. These percentages point to children's small but perceptible contribution to pulling households out of poverty, that contribution being somewhat larger in Pakistan. Relative to Peru, the proportion of children in Pakistan who have never at- tended school is alarmingly high. Further, there is a marked gender disparity in the educational experience of adults in Pakistan. On average, the most educated woman in a Pakistani household has only 39 percent of the schooling of the most educated man. This figure is 89 percent in Peru. The educational deprivation of women in Pakistan is further reflected by the fact that, unlike in Peru, women have received much less schooling than their children. Relative to a man, a woman works fewer hours in outside, paid employment in Pakistan. The gap is much smaller in Peru. This feature, which reflects Islamic Table 3. Household Characteristics of Peru and Pakistan (sample means) Characteristic Peru Pakistan Number of children in the household 3.84 5.61 Ratio of girls to boys 0.51 0.48 Percentage of children living in households below the poverty line* Boys 30.42 27.11 Girls 29.37 25.57 Ratio of the most educated woman's educational experience to that of the most educated man in the household11 0.89 0.39 Ratio of the child's to the man's educational experience1' 0.64 0.68 Ratio of the child's to the woman's educational experience1' 0.72 1.75 Average age of child 11.38 13.16 Percentage of households that are female headed 13.17 1.87 Percentage of children living in households with electricity 61.33 76.45 Percentage of children living in urban areas 59.50 53.67 Ratio of child's labor hours to man's labor hours 0.12 0.13 Ratio of child's labor hours to woman's labor hours 0.25 0.60 Percentage of children involved in child labor 26.27 23.36 Percentage of children who have not received any schooling 2.00 31.92 a. The poverty line is set at 50 percent of the median nonchild household income per adult equivalent in the sample. b. Measured in years of schooling. Source: 1994 Peru Living Standards Measurement Survey and 1991 Pakistan Integrated Household Survey. Ray 353 laws that constrain women's involvement with outside work, is also seen in the much higher ratio of children's labor hours to women's labor hours in Pakistan than in Peru. Peruvian women contribute a much larger share of household earnings than do Pakistani women (table 4). This is consistent with the sample characteristics presented in table 3, especially the fact that the difference in educational levels between Pakistani women and Pakistani men is much greater than the difference between Peruvian women and Peruvian men. Hence, Pakistani women have much less earning power than Pakistani men. Moreover, households in Pakistan are much more dependent on children's earnings than are households in Peru. Chil- dren in Pakistan lag only marginally behind women in their share of household earnings. This largely explains the result, presented and discussed later, that com- mercially paid hours of child labor are much more responsive to household pov- erty in Pakistan than in Peru. n. ESTIMATION METHOD AND EMPIRICAL RESULTS The empirical method is based on the two-step procedure, discussed in Maddala (1983), for estimating labor supply equations after correcting for sample selec- tivity. The results were obtained using the liMDEP program written by Greene (1995). I ensure that the estimated child labor and child schooling equations are pure reduced-form equations with none of the variables on the right side likely to suffer from endogeneity. Testing the Luxury Axiom Since a key motivation of this exercise is to test Basu and Van's (1998) luxury axiom, I define the poverty status of a household with respect to a poverty line set at 50 percent of the median nonchild household income per adult equivalent of the sample. Peru and Pakistan disagree on the validity of the luxury axiom (table 5). The estimated coefficient of the poverty variable is weak and statisti- Table 4. Shares of Household Earnings in Peru and Pakistan (sample means) Household member Peru Pakistan Man's share 0.745 0.853 (0.325) (0.251) Woman's share 0.239 0.088 (0.319) (0.193) Child's share 0.016 0.059 (0.076) (0.161) Note: The (ample consists of 2,873 households in Peru and 3,720 households in Pakistan. Figures in parentheses denote standard deviations. Source: 1994 Peru Living Standards Measurement Survey and 1991 Pakistan Integrated Household Survey. 3S4 THE WORLD BANK ECONOMIC REVIEW, VOL 14, NO. I Table 5. Regression Estimates of Child Labor Supply Equations Coefficient estimate 1 Peru Pakistan Variable Boys Girls Boys Girls Constant -3,439.6* -2,847.0* -8,769.40* -9,986.00* (448.65) (508.84) (2,108.08) (2,734.75) Child characteristics Age 289.49* 288.79* 883.09* 732.07 (71.81) (81.45) (311.79) (411.55) Age2 -3.38 -6.31 -18.02 -21.17 (2.99) (3.43) (11.45) (15.35) Wage 484.92* 433.50* 87.88* 138.05* (52.32) (64.32) (10.80) (23.60) Family characteristics Poverty status (1 if below poverty line, 0 otherwise) -58.63 -92.92 491.63* 472.51* (76.68) (84.65) (125.91) (172.85) Region of residence (1 = urban, 0 = rural) -843.28* -992.09* -480.75* -288.99 (93.16) (106.46) (136.56) (188.69) Number of children 50.93* 13.92 -11.75 -35.30 (17.21) (20.73) (20.13) (27.13) Number of adults 12.56 -29.27 -46.68 -57.91 (27.60) (31.67) (30.50) (39.08) Gender of household head (0 = male, 1 = female) -15.19 72.72 -221.34 -511.90 (100.96) (115.04) (361.19) (509.51) Age of household head 1.62 3.34 0.50 16.32* (3.17) (3.42) (4.47) (6.00) Man's education -21.70* -17.09 -74.42* -75.44* (8.06) (926) (12.21) (16.61) Woman's education -22.10* -143.1** -77.41* -96.94* (8.29) (10.08) (17.12) (25.61) Man's wage* -43.97 -112.42* 5.77 -11.28 (38.08) (28.49) (10.10) (14.06) Woman's wage1 -16.12 -18.13 1.85 124.56* (28.45) (27.29) (14.97) (18.66) (Man's wage)2 -2.08 1.74* -0.17 0.20 (3.41) (0.60) (0.14) (OJZO) (Woman's wage)2 . 0.69 0.21 0.02 -1.90* (0.82) (0.53) (0.26) (0.40) Community characteristics Water storage (1 o best, 6 = worst) 36.76 -70.39* 113.57* 23.38 (19.31) (21.97) (30.41) (39.43) Disposal of sewage (1 e best, 6 = worst) 111.03* 119.52* -286.14* 348.57* (22.51) (25.64) (53.75) (76.08) Electricity (1 o yes, 0 = no) -198.86* -151.80 235.34 679.71** (9027) (101.62) (233.59) (317.03) Ray 35S Table 5. (continued) Coefficient estimate • l Peru Pakistan Variable Boys Girls Boys Girls Joint tests Community variables %\ = 47.39* = 28.52* x\ x\ = 34.53 * x\ = 26.75* •Significant at the 1 percent level. * * Significant at the 5 percent level. Note: Standard errors are in parentheses. a. In the case of households with more than one working man or more than one working woman, I take the mnyimnm wage earned as a measure of the man's and woman's wage, respectively. Source: 1994 Peru living Standards Measurement Survey and 1991 Pakistan Integrated Household Survey. cally insignificant for both girls and boys in Peru.7 The reverse is true for Paki- stan. A Pakistani household that was previously not poor will increase the out- side, paid employment of its children substantially, by approximately 500 child labor hours annually for each child, if it falls below the poverty line—exactly as predicted by the luxury axiom. This mixed response is consistent with our earlier observation that Pakistani children, especially boys, play a greater role in pulling households out of poverty than Peruvian children. The Pakistani result appears to contradict Bhatty (1998: 1734), who cites a variety of empirical studies on Indian child labor in support of the view that "income and related variables do not seem to have any direct significant effect on children's work input." However, as I report below, this contradiction is partly resolved by extending the definition of child labor to include domestic work. The evidence in favor of the luxury axiom weakens for Pakistani girls. Testing the Significance of Wages and Other Explanatory Variables It is possible to make other interesting comparisons between Peru and Paki- stan. In both countries the coefficient of the age-squared variable is insignificant, suggesting a linear relationship between children's labor hours and age. In Paki- stan increasing age has a greater impact on boys' labor supply than on girls', whereas in Peru there is no gender differential. However, this differential may be more apparent than real, since older girls in Pakistan are likely to spend time in unpaid domestic work, which we have not yet considered in the child labor estimations. Ceteris paribus, in Peru and Pakistan both boys and girls living in urban areas work fewer hours than their rural counterparts, possibly reflecting the impor- tance of farm employment as a destination of child labor in these primarily agri- cultural countries. However, in absolute terms, whereas the estimated coefficient of the regional dummy variable for girls' labor exceeds that for boys' labor in Peru, the reverse is indicated for Pakistan. 7. In response to a referee's suggestion that the statistical insignificance of the poverty coefficient in the case of Peru reflects mulrkolinearity between the poverty and the wage variables, I dropped all of the wage variables and still found the estimated poverty coefficient to be statistically imrignifimnt. 356 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 Although both countries agree on the positive role that increasing the educa- tion of adults in the household plays in reducing both boys' and girls' labor, the size and significance of these effects are much stronger in Pakistan than in Peru. Moreover, most of the community variables have a significant impact on hours of child labor, with most of the estimated coefficients having the expected sign. For example, a deterioration in the sewage disposal system sharply increases girls' labor hours in Pakistan and boys' and girls' labor hours in Peru, but it reduces boys' labor hours in Pakistan. The significant impact of community variables on child labor is confirmed by the rejection of the joint hypothesis of no community infrastructure effects (table 5). Without exception, hours of child labor respond positively and significantly to child wages. Of particular interest are the estimated coefficients of the adult wage variables, which indicate the nature of the interaction between the child and adult labor markets. The estimated coefficients differ in sign and magnitude be- tween girls and boys and between countries. In Peru increasing adult wages re- duce boys' and girls' labor hours, suggesting that child and adult labor hours are substitutes in that country. In Pakistan rising women's wages sharply increase the labor hours of girls but have a negligible impact on the labor hours of boys. The statistically significant and negative coefficient on the quadratic of women's wages in the case of Pakistani girls suggests that this relationship is nonlinear and weakens in the higher age categories (see also Basu 1993). Thus in Pakistan, without good schools and satisfactory day care arrangements, mothers who work have to put their children to work as well. Comparing the estimated coefficients of the adult wage variables for Pakistan shows that the interaction between the children's labor market, especially that of girls, and the women's labor market is different from that between the children's and the men's labor markets. In response to a referee's suggestion, I reestimate the child labor equations for Pakistan using the woman's employment status as a dummy variable (1 if she works, 0 otherwise), in place of the woman's wage variable (given in table 5). 8 The results, which are presented in table 6, show that the estimated coefficient of the woman's employment status is positive and statis- tically significant for both boys and girls. In other words, children from house- holds in which a woman is employed work longer hours than other children. This is consistent with the evidence presented in table 5 and confirms the close complementarity between girls' and women's labor markets in Pakistan. More- over, the size and significance of the estimated coefficient of the woman's em- ployment status is much higher for girls than for boys, suggesting a stronger relationship between women's and girls' labor markets. Including Domestic Work The results presented in table 5 relate to children in full-time, paid employ- ment. To investigate the robustness of these results to a more relaxed treatment 8. In most cases the woman is the child's mother, but ihe may be the child's elder sister or aunt. Ray 357 Table 6. Regression Estimates of Child Labor Supply Equations in Pakistan using Woman's Employment Status as a Determinant Coefficient estimate Variable Boys Girls Constant -9,095.1* -1,1733.0* (2,114.8) (2,603.9) Child characteristics Age 904.2* 866.4** (312.5) (391.0) Age 2 -18.7 -26.1 (11-5) (14.6) Wage 86.5* 98.1* (10.8) (19.6) Family characteristics Poverty status (1 if below poverty line, 0 otherwise) 520.0* 487.0* (126.2) (163.6) Region of residence (1 = urban, 0 = rural) -414.6* 118.7 (138.4) (183.9) Number of children -11.6 -32.8 (20.2) (26.1) Number of adults -51.3 -105.0* (30.6) (38.4) Gender of household head (0 = male, 1 « female) -215.0 -533.2 (3603) (487.4) Age of household head 0.0 15.4* (4.5) (5.7) Man's education -69.8* -40.1" (12.2) (15.9) Woman's education -79.3* -98.4* (17.1) (24.6) Man's wage* 8.0 3.1 (10.1) (13.4) Woman's employment status (1 if woman works, 0 otherwise) 336.9* 2^576.9* (107.1) (138.1) (Man's wage)2 -0.2 0.1 (0.1) (0.2) Community characteristics Water storage (1 = best, 6 = worst) 114.5* 45.2 (30.5) (37.9) Disposal of sewage (1 = best, 6 = worst) -288.8* 288.3* (53.8) (72.8) Electricity (1 = yes, 0 = no) 260.7 572.3 (234.2) (302.5) * Significant at the 1 percent level. * * Significant at the 5 percent level. Note: Standard errors are in parentheses. a. In the case of households with more than one working man, I take the maYimnm wage earned as a measure of the man's wage. Source: 1991 Pakistan Integrated Household Survey. 358 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 of child labor, I now go beyond the HO definition of child labor to include hours that children spend on domestic work. This information was, however, only avail- able for Pakistani girls. Therefore I estimate labor equations for Pakistani girls with and without domestic duties (table 7). Including domestic duties in the definition of child labor weakens the impact of poverty on a girl's labor hours, which fall from 473 to 60 a year. The statisti- cal significance of the estimated coefficient of the poverty variable also disap- pears. In other words, Pakistani girls fall in line with Peruvian children in failing to support the luxury axiom. Clearly, when a household falls into poverty in Pakistan, girls, having to work more hours in outside, paid employment, are able to devote fewer hours to household chores. Consequently, the overall impact of poverty on their labor hours weakens sharply. The absolute magnitude and significance of several other variables also weaken sharply. Rising woman's education continues to have a significant negative im- pact on hours of child labor. However, the impact of a man's education lessens and loses its statistical significance when domestic duties are included. Similarly, the magnitude and significance of the woman's wage coefficients weaken consid- erably, with the coefficient on squared wages losing its strong statistical signifi- cance. However, the different response of child labor hours in Pakistan to changes in a man's and woman's wages is robust to the alternative definitions of child labor. On a likelihood ratio test the community variables continue to have a significant impact on child labor, even in the presence of domestic duties. Does the Luxury Axiom Hold for Schooling? If we extend the luxury axiom to the context of schooling, a household would not send its children to school if it falls into poverty. A negative and statistically significant coefficient on the poverty variable would confirm this axiom. As in the case of child labor, the Peruvian evidence on child schooling, especially school- ing of girls, contradicts the luxury axiom, and the Pakistani evidence confirms it (table 8). The results for Pakistan hold for both sexes, especially girls. This confirms the earlier observation, also noted by Weiner (1996), Basu (1999), and others, that South Asian children, especially girls from poor households, drop out of school to enter the labor market. The lack of good schools in Pakistan, along with the consequent discount that parents place on the value of their children's education, may explain this behavior. The gender differential in this respect is quite reveal- ing, with Pakistani girls experiencing a much sharper reduction in their schooling than boys when their households fall into poverty. Relative to Pakistani children, Peruvian children register much smaller and statistically insignificant changes in their schooling as they slip into poverty, with Peruvian girls even registering a small increase. This reflects partly the superior educational experience of Peruvian children in relation to children in Pakistan and partly the fact, as observed by Patrinos and Psacharopulos (1997) in their study of child labor in Peru, that Peruvian children combine employment with Ray 359 Table 7. Sensitivity of Regression Estimates of Girls' Labor Supply Equation in Pakistan to the Inclusion of Domestic Work . , Coefficient estimate Domestic hours Domestic hours Variable included excluded Constant -3,550.20* -9,986.00* (1,236.79) (2,734.75) Child characteristics Age 581.32* 732.07 (185.56) (411.55) Age* -13.98" -21.17 (6.98) (15.35) Wage 5628* 138.05* (14.89) (23.60) Family characteristics Poverty status (1 if below poverty line, 0 otherwise) 59.55 472.51* (83.88) (172.85) Region of residence (1 = urban, 0 = rural) -254.28* -288.99 (86.68) (188.69) Number of children -8.33 -35.30 (12.18) (27.13) Number of adults -107.54* -57.91 (16.93) (39.08) Gender of household head (0 = male, 1 m female) -304.15 -511.90 (232.64) (509.51) Age of household head 2.70 16.32* (2.8J) (6.00) Man's education -14.08 -75.44* (7.26) (16.61) Woman's education -57.60* -96.94* (9.47) (25.61) Man's wage* -3.78 -11.28 (6.00) (14.06) Woman's wage* 20.24** 124.56* (9.38) (18.66) (Man's wage)2 0.07 0.20 (0.08) (0.20) {Woman's wage)2 -0.32 -1.90* (0.21) (0.40) Community characteristics Water storage (1 =• best, 6 = worst) 7.39 23.38 (18.52) (39.43) Disposal of sewage (1 = best, 6 = worst) 83.56** 348.57* (33.35) (76.08) Electricity (1 = yes, 0 = no) -193.56 679.71 *» (151.85) (317.03) Joint tests Community variables %\= 10.39** X 2 -26.75* * Significant at the 1 percent level. ** Significant at the 5 percent level. Note; Standard errors are in parentheses. a. In the case of households with more than one working man or more than one working woman, I take the ma-rimnm wage earned as a measure of the man's and woman's wage, respectively. Source: 1991 Pakistan Integrated Household Survey. 360 THE WORLD BANK ECONOMIC REVIEW, VOL 14, NO. 2 Table 8. Regression Estimates of Child Schooling Equations . 1 Coefficient estimate Peru Pakistan Variable Boys Girls Boys Girls Constant -5.38* -5.99* -7.55* -11.19* (0.88) (0.95) (2-54) (3.57) Child characteristics Age 0.98* 1.08* 0.975** 1.690* (0.15) (0.15) (0.38) (0.54) Age* -0.011 -0.018* -0.018 -0.056* (0.01) (0.01) (0.014) (0.020) Wage -0.204 -0.122 -0.194* -0.166* (0.13) (0.17) (0.023) (0.055) Family characteristics Poverty status (1 if below poverty line, 0 otherwise) -0.175 0.220 -0.476* A.J.65* (0.17) (0.17) (0.169) (0.250) Region of residence (1 = urban, 0 = rural) 0.018 -0.023 -0.258 0.662* (0.21) (0.21) (0.176) (0.251) Number of children -0.095** -0.098** 0.052** 0.010 (0.04) (0.04) (0.025) (0.035) Number of adults -0.162* -0.068 -0.174* -0304* (0.06) (0.06) (0.037) (0.049) Gender of household head (0 = male, 1 = female) -0.165 0.083 0.828 1.954* (0.21) (0.21) (0.473) (0.630) Age of household head 0.007 0.005 0.005 0.009 (0.01) (0.01) (0.006) (0.008) Man's education 0.040** 0.050* 0.259* 0.296* (0.02) (0.02) (0.015) (0.021) Woman's education 0.047* 0.049* 0.152* 0.367* (0.02) (0.02) (0.019) (0.025) Man's wage* 0.005 0.002 0.014 0.015 (0.03) (0.04) (0.012) (0.017) Woman's wage* 0.003 0.046 0.015 -0.072* (0.05) (0.04) (0.020) (0.028) (Man's wage)2 0.000012 -0.00016 0.000 0.000 (0.00) (0.00) (0.000) (0.000) (Woman's wage)1 0.00029 -0.0010 -0.001 0.001 (0.00) (0.00) (0.00) (0.00) Community characteristics Water storage (1 = best, 6 = worst) -0.027 -0.047 -0.097* * -0.188* (0.04) (0.05) (0.039) (0.055) Disposal of sewage (1 = best, 6 = worst) -0.066 -0.075 0.062 -0.596* (0.05) (0.05) (0.069) (0.094) Electricity (1 = yes, 0 = no) 0327 0.517** 1.002* 0.668 1021) (0.21) (0.313) (0.494) Ray 361 Table 8. (continued) Coefficient estimate I Peru Pakistan Variable Boys Girls Boys Girls Joint tests Community variables X5-7.12 X\ = 14.82* xi = 19-60* X\ = 75.86* 'Significant at the 1 percent level. * * Significant at the 5 percent leveL Note: Standard errors are in parentheses. a. In the case of households with more than one working man or more than one working woman, I take the maYimnm wage earned as a measure of the man's and woman's wage, respectively. Source: 1994 Peru Living Standards Measurement Survey and 1991 Pakistan Integrated Household Survey. schooling to a greater extent than in other countries. Patrinos and Psacharopulos (1997:398) comment, "Working actually makes it possible for the children to go to school." Raising the education of adults has a positive effect on children's schooling in both countries. Parents who are more educated, especially those in Pakistan, are better able to see the value of their children's education and to resist the tempta- tion to pull them out of school. Rising child wages increase the opportunity cost of education and significantly reduce child schooling in Pakistan. But the effect in Peru is statistically insignificant. The close complementarity between girls' and women's labor in Pakistan is consistent with the negative impact that rising women's wages have on child schooling. In other words, when women's wages rise, working mothers tend to pull daughters out of school and take them along to work. There is no such relationship in Peru. It is also interesting to observe that girls in Pakistan, although not boys, experience significantly more schooling in urban areas than in the rural countryside. Estimating Earnings Share Equations The regression estimates presented and discussed in table 5 relate to the indi- vidual labor supply behavior of children in the household. This is in line with the collective models of household behavior that have been proposed recently (see, for example, Alderman and others 1995). Although such models are attractive from a policy viewpoint, especially since welfare analysis stems from and should explicitly relate to individual behavior, they have one significant limitation in the current context: decisions about a child's labor force participation and hours of work, leisure, and schooling are typically made by the adult, not by the child. The conventional treatment of the household in unitary models may therefore be relevant. Moreover, the share of household earnings contributed by men, women, and children and their responsiveness to changes in key household and commu- nity characteristics are of direct policy concern. The following approach main- tains the assumption of joint welfare maximization underlying the unitary model. 3 62 THE WORLD BANK ECONOMIC REVIEW, VOL. 14, NO. 2 To provide empirical evidence on this issue, I estimate the following three- equation pystem of earnings shares: (1) W; =