The Effect of Domestic Work on Girls' Schooling: Evidence from Egypt

Abstract

In Egypt, girls' work primarily takes the form of domestic tasks, which are not considered in many studies of child labor. This paper investigates the effect of girls' work on their school attendance. It uses a modified bivariate probit approach to estimate the effect of work on schooling while allowing for the simultaneous determination of the two outcomes. It presents evidence that the substantial burden of girls' domestic work leads to lower rates of school attendance. Policies that attempt to ban the labor-force work of children will have practically no effect on girls' education in Egypt, while interventions reducing the drudgery of household labor through, for example, improved water and sanitation infrastructure, have better prospects for success.

Keywords:

• Child labor
• schooling
• domestic work
• gender
• Egypt
• household economics

Acknowledgments

We thank Cristobal Ridao-Cano for allowing us to adapt his Stata program for a switching probit model. We are grateful to Dennis Ahlburg, Ray Langsten, Brian McCall, David Post, and Insan Tunali for detailed and valuable comments. Sara Kaufman, Pingkang Yu, Yi Zheng, and Hongliang Zhang provided competent research assistance. We gratefully acknowledge financial assistance and technical support from the International Center for Economic Growth (ICEG) through its Economic Policy Initiative Consortium project in Egypt, which is funded by USAID.

Notes

The possible exception is the literature on the effect of mother's participation in the labor market on girl children's education, which implies that mothers' and girls' domestic work are potential substitutes (for example, Lourdes Benería [1992] and Deborah Levison, Karine S. Moe, and Felicia Knaul [2001]).

These figures are obtained from the 1988 Labor Force Sample Survey (LFSS) and the 1998 Egypt Labor Market Survey (ELMS). The labor-force participation rates reported here are based on an extended definition of economic activity and a short reference period of one week with a one-hour-per-week minimum threshold. The 1988 survey did not measure the incidence of domestic work.

This assumption also has its problems. When schools are of poor quality, children may benefit more from other activities. When children are regularly beaten and verbally abused in schools, our assumption is again problematic.

Examples include Deborah S. DeGraff and Richard Bilsborrow (2003) for the Philippines, Peter Jensen and Helena Skyt Nielsen (1997) for Zambia, Sudharshan Canagarajah and Harold Coulombe (1998) for Ghana, Felicia Knaul (1995) for Mexico and Colombia, Melissa Binder and David Scrogin (1999) for Mexico, Harry Anthony Patrinos and George Psacharopoulos (1997) for Peru, George Psacharopoulos (1997) for Bolivia and Venezuela, George Psacharopoulos and Ana Maria Arriagada (1989) for Brazil, Jose Canals-Cerda and Cristobal Ridao-Cano (2004) for Bangladesh, and Emmanuel Skoufias (1994) for India.

Canals-Cerda and Ridao-Cano (2004) use a similar approach to determine the effect of working while in school on the probability of progressing to secondary school in rural Bangladesh.

Ragui Assaad was principal investigator, responsible for the sampling design, instrument, training of enumerators, and implementing of the survey.

The sixth year was phased back in starting with the children who entered the first grade in 2000.

For an overview of child labor in Egypt, see Zibani (2002) and Kawther Abu Gazaleh, Lamia Bulbul, Soheir Hewala, and Suadad Najim (2004).

Similarly, Marcel Fafchamps and Jackline Wahba (2006) examine three types of work, including market work, subsistence work, and household chores.

These figures are based on a one-week reference period with a one-hour minimum threshold. The figure is 2 million (32 percent) if a fourteen-hour-per-week threshold is used instead.

Because of the way the questionnaire is designed, we do not observe the number of hours in subsistence work for girls who are engaged in market work; thus, our inclusive measure will understate the total hours of work for the small number of girls engaged in both market work and subsistence work.

Our reasoning is not completely ad hoc. Out of twenty-four hours, if a child sleeps for nine hours, spends two hours on meals and self-care, six hours in school, one hour going to and from school, one hour studying, 0.5 hours in tutoring sessions (about 53 percent of in-school 6–14-year-olds were tutored in 2006), and two hours in various kinds of work activities, that leaves only 2.5 hours for TV-watching, play, and other social activities. William Kandel and David Post (2003) also used a fourteen-hour cut-off, with similar justifications.

The slight dip in school attendance at age 10 is most likely due to age heaping. Lack of precision about age (and, thus, heaping on 10) is more likely for those children who do not go to school because parents of children who are out of school are likely to be illiterate and therefore not sure of the ages of their children.

We include a cluster correction to take into account that these variables are not independent across observations in the same primary sampling unit (PSU). There are 200 PSUs in the sample. A village or neighborhood was typically represented by at most one PSU, with the exception of one or two that had two PSUs.

Irregular private-sector work is the omitted category. “Regular private sector” jobs consist of permanent and temporary but continuous jobs in the private sector, while “irregular private sector” jobs consist of intermittent and seasonal jobs. Public-sector work is typically regular. Nonwage workers are either employers, self-employed workers, or, in some rare cases, unpaid workers for a family enterprise. Non-working fathers are either unemployed or out of the labor force. We expect that fathers in some types of positions are more likely to be able or willing to take their sons (but not daughters) to work with them. Nonworking fathers may stay home and generate more household work for daughters.

We obtained the following test statistics for a Wald test of the equality of the two correlation coefficients: Model 1 had χ²(1) = 0.05, p-value = 0.821; Model 2 had χ²(1) = 0.07, p-value = 0.793; and the unrestricted Model 3 did not converge. An insignificant test statistic upheld the restriction.

Victor Levy (1985) relies on changes in cropping patterns to identify the effect of child labor in agriculture on the total fertility rate in Egypt.

According to the 1998 ELMS, only 6.3 percent of adults who ever worked changed their place of residence in the ten years prior to the survey.

Because work is endogenous to the schooling decision, some value has to be set for the probability of work to calculate the marginal probability of schooling and the joint probabilities of work and school. We use a value of 0.46, the unconditional probability of work of the reference girl. For the other characteristics of the reference girl, see note to Table 4.

A more detailed discussion of the control variables can be found in Ragui Assaad, Deborah Levison, and Nadia Zibani (2007).

In the absence of controls for the father's income, the father's education variable may partially capture the effect of income on a girl's work and schooling. We address this to some extent by including the wealth quintiles, which should be highly correlated with family income.

Eva Mueller (1984) documents that rural children in Botswana are more likely to work if their families are wealthy enough to own complementary assets, such as land, farming implements, and livestock.

Brian McCall (1992) shows that a binary choice model with a binary endogenous regressor is just identified without exclusion restrictions if the underlying dependent variable is linearly related to the explanatory variables and if two of the explanatory variables are continuous. The error terms are also assumed to be continuous. Lung Fei Lee (1992) shows that this over-identification test statistic is distributed as χ² with degrees of freedom equal to the number of excluded instruments.

The over-identification test has five degrees of freedom, one for each of the five instruments. We obtain the following test statistics: Model 1 had χ²(5) = 3.49, p-value = 0.625; Model 2 had χ²(5) = 3.65, p-value = 0.601; and Model 3 had χ²(5) = 4.07, p-value = 0.539. Thus, the results are as we had hoped, and exclusion of the instruments from the schooling equation is upheld. Since the likelihood ratio test does not allow for a cluster correction, this test was carried out without it.

We obtain the following test statistics for a Wald test of joint significance: Model 1 had χ²(5) = 28.46, p-value = 0.0001; Model 2 had χ²(5) = 25.83, p-value = 0.0001; and Model 3 had χ²(5) = 22.33, p-value = 0.0005.