Factors influencing ‘missing girls’ in South Korea

Abstract

Despite the massive attention drawn to ‘missing girls,’ there has been no study that specifically focuses on the association between childlessness and the daughter deficit. Using a bivariate probit selection model, this article analysed the data for 6475 married women aged 15–49 years collected from the 2003 Korea National Fertility and Family Health Survey. The results showed that a couple's decision to have a child exerted a significant influence on its daughter deficit. This study also found that the effect of a woman's education on daughter deficit did not correspond to that of her husband's level of education. Additionally, a prediction was made that if a one child family norm were prevailing in South Korea, the probability of a couple's having a daughter deficit would increase by as much as 63.9%.

Acknowledgements

We are grateful to an anonymous referee for making comments on the identification restriction. The findings, interpretations and conclusions expressed in this article are entirely those of the authors. They do not necessarily represent the view of the World Bank, its executive directors or the countries they represent.

Notes

¹ We had considered a variety of measures of the size of the daughter deficit. For example, we took into account the sex ratio, which is defined as the ratio of males to females. However, this measure cannot capture the cases where the number of daughters is zero. As another example, the excess number of sons (the number of sons minus the number of daughters) was considered. This was also not appropriate as a dependent variable in ordinary least squares regression model because the range of the variable was very short. Thus, we tried an ordered logistic model, but found that it violated the proportionality of odds. Finally, we came to the conclusion that the dummy variable indicating whether a woman has a daughter deficit is the best fit for the purposes of this study.

² Although the more common estimation procedure is the Heckman two-step estimation, the major reasons that we decided to employ the bivariate probit model with sample selection are as follows. First, Heckman's estimator can be affected by multicollinearity between the estimated Mill's ratio and explanatory variables (Nawata, 1993, 1994; Nawata and Nagase, 1996). Second, it seemed inappropriate to use the least squares regression in the second step of Heckman's procedure because the number of children (sons or daughters) was not well distributed enough to assume that it is a continuous dependent variable. To overcome this, we also conducted ordinal level models such as ordered logit and probit but the model did not satisfy the conditions of the proportionality of odds.

³ The reasons that we selected woman's age at marriage as a variable instead of woman's age as a variable are as follows. First, many previous studies have suggested that a woman's age at marriage is an important factor in determining the daughter deficit. As presented in the manuscript, women who conform less often to tradition are likely to marry at an older age (Larsen et al., 1998; Hussain et al., 2000) and likely to have a weaker daughter deficit. Second, chi-squared analyses showed that the relationship between the woman's age at marriage variable and the daughter deficit variable (p < 0.0001) was more statistically significant than that between the woman's age variable and the daughter deficit variable (p = 0.0192). Third, severe multicollinearity was detected either (i) in the case where both the woman's age variable and the woman's age at marriage variables were included in multivariate analyses (maximum value of the VIF = 21.9143) or (ii) in the case where only the woman's age variable was included (maximum value of VIF = 21.3417). In contrast, when we only included the woman's age at marriage variable, the value of VIF was very low (maximum value of VIF = 1.9115). Allison (1999) suggested that a VIF greater than 2.5 is problematic when using linear regression models to measure multicollinearity in a dichotomous outcome analysis. Thus, the multicollinearity was not a problem in the case where only the woman's age at marriage variable was included.

⁴ Data were analysed using SAS statistical software version 9.1 (SAS Inc., Cary, NC) and LIMDEP statistical software version 8.0 (Econometric Software Inc., Plainview, NY). Analyses were conducted using p < 0.1 as the significance level.

⁵ The method is available upon request from the first author.

⁶ In order to find an explanatory variable, which satisfies the identification restriction (Reize, 2001, p. 3; Averett et al., 2002, p. 1774 and Table 2; Sartori, 2003, p. 112; van der Klaauw and Koning, 2003, p. 32), we examined all explanatory variables in the model and found the ONLY variable (indicating whether a woman's husband is an only son) appropriate for the identification restriction, both because it affected childlessness significantly (p = 0.0258) but not the daughter deficit (p = 0.6663) and because it seems less important than other explanatory variables like woman's level of education.