The labour market impact of body weight in China: a semiparametric analysis

Abstract

While a positive wage effect of Body Mass Index (BMI) is widely observed in low-income developing countries, a negative wage effect of BMI is often observed in high-income developed countries. To fill the gap between these previous findings, we investigate the relationship between body weight and wages in transition economies. We focus on China, whose rapid economic growth of the 1990s was followed by a rapid increase in overweight and obesity while still experiencing significant food insecurity and underweight. we first use several parametric regression strategies to obtain a consistent estimate of the wage effects of weight. Second, we adopt a semiparametric partially linear model that allows for endogeneity of weight. Parametric regressions provide mixed results, and the sign and magnitude of their estimates are sensitive to the choice of samples and regression strategies. Semiparametric estimates provide evidence of a wage penalty for very heavy and thin persons among both men and women. The wage penalty is more significant among men than among women. Semiparametric results also indicate that parametric estimates can overstate and misrepresent the wage effects of weight for healthy weight persons due to their restrictive functional form assumptions.

Acknowledgements

The author would like to thank an anonymous referee and Mark Taylor for valuable comments on an earlier version of this article. The author would also like to thank John Cawley, Per Pinstrup-Andersen and Kiran Gajwani for helpful suggestions.

Notes

¹ WHO (1995) plays a key role in establishing a consensus on how to interpret anthropometric indicators in the context of health and nutritional status.

² Strauss (1986) first described the endogeneity problem of nutritional status in econometric terms and applies the method of IVs to a labour market in Sierra Leon. Similarly, Sahn and Alderman (1988) examine a labour market in Sri Lanka controlling for endogeneity. However, these studies examine only the effect of nutrient intake, and thus are not referred to specifically in this section.

³ In our sample, 2.6% of male workers and 2.1% of female workers are either under 18 years or over 60 years and are excluded.

⁴ We use the 1st and 99th percentiles of the union of our two CHNS samples to obtain the same domain of BMI in the two samples. Ninety-nine and fifty eight observations are dropped from the men and the women sample, respectively.

⁵ World Health Organization (WHO, 1995) recommends to use the 5th, 85th and 95th percentiles in Must et al. (1991) to define underweight, overweight, and obesity for people aged 18–24 years, respectively. For example, the cut-off values for overweight range from 25.5 to 26.9 for people aged 18–24 years. However, for the sake of parsimony, we adopt the same cut-off values for the whole sample.

⁶ See Baum and Ford (2004) for more discussions about the channels through which nutritional status affects wages. Madden (2004) also attempts to identify labour market discrimination due to poor health conditions in the United Kingdom.

⁷ Averett and Korenman (1996) combined the family fixed effects model with the lagged measure of weight.

⁸ Although a more desirable instrument is the weight measure of siblings who live separately (see, e.g., Cawley, 2004), the data on a sibling relationship across households are not available in the CHNS data.

⁹ This does not mean that family environment has no effect on the body weight of household members. This means that the variance in the weight of household members which is explained by shared family environment is mostly explained by genetic factors. Thus, once genetic factors are controlled, there remains little variance in body weight that is explained by shared family environment. On the other hand, the family environment that is not shared among household members is reported as an important determinant of the nongenetic variance in weight.

¹⁰ The bandwidth is determined by the Sheather–Jones plug-in method. The kernel function used is Gaussian.

¹¹ Using the occupation types defined in the CHNS, white-collar workers are senior professional/technical workers, junior professional/technical workers, administrator/executive/managers, office staff, army officers, police officers, service workers. Blue collar workers are farmers, fishermen, hunters, skilled workers, nonskilled workers, soldiers and drivers.

¹² The coefficient on the IMR was statistically significant at the 10% level in 8 and 6 regressions out of 29 regressions in the male and female group, respectively.

¹³ Regression results for other covariates are available from authors upon request.

¹⁴ The proportion of the population in which the indicator for underweight, overweight and obesity changed is 6.55, 10.7, and 1.13%, respectively.