Duke 2008 | ASHE | Equity and Efficiency in Health and Healthcare

The association between education and health is well established. Whether or not this association reflects a causal effect of education on health is less established, though. Education may proxy for unobserved individual characteristics, such as family background and genetic traits, in which case an estimated coefficient of education on health is upward biased. In this paper, my aim is to estimate the causal effect of education on health and health-related behaviors. For this purpose, I will use nationally representative data, including almost 2,000 twins, from the MIDUS survey in the US. Twins share common background and, in the case of identical twins, common genes. By running regressions on within-twin-pair difference, I am able to purge out the influence of omitted variable, such as genetic characteristics and parental background. Any remaining correlation is more likely to reflect a causal effect. The resulting estimates will then be compared to estimates using the non-twin main sample. To the best of my knowledge, this is the first study using a twin design to study the effect of education and health. Recent studies use educational reforms to identify the effect of education on health (Arendt 2005; Lleras-Muney 2005). These studies estimate local average treatment effects and the effects are thus only identified for the subgroup that is affected by the reform. Estimates based on a twin strategy overcome this important shortcoming. Using twins, I am able to study the effects of education on health on the entire distribution. A twin research strategy therefore has the potential to provide a major advancement in our understanding on the effect of education on health. Twin studies hinges on the assumption that unobserved ability that determine schooling is cancelled out when using twin fixed effects. Previous studies in labor economics have only to a limited extent been able to address whether there exist remaining unobserved difference between twins after such differencing that will still determine schooling (Ashenfelter & Kreuger 1994). It has been argued, for instance, that difference in birth weight and difference in treatment by parents may determine schooling outcomes (Bound & Solon 1999). The MIDUS data contains information on such factors and I am thus able to analyze whether these factors are important determinants of within-twin difference in schooling. difference in education between twins may also stem from difference in health during early life. While controlling for birth weight picks up some of these difference, difference in health may also appear later during adolescence. To rule out such reverse causality I control for self-assessed health during adolescence. Preliminary results show that the effect of education on health and health behaviors, such as smoking, is substantially reduced when using twin fixed effects compared to OLS results. This is what one would expect if twin differencing reduces the influence of unobserved genetic traits and family background. The difference between the twin and OLS estimate is larger than one would expect given measurement errors in education that are exaggerated when taking twin difference and, thus, downward biasing the twin estimates.