The Relationship between Education and Health Behaviors: Is it Causal?
- Presenter:
Chair: Glenn Blomquist; Discussant: Judith Shinogle Tue June 6, 2006 8:00-9:30 Room 325
An extensive body of literature documents a relationship between formal education and health that is strong, broad, and persistent. When presented with such a robust association, it is natural to make the leap, implicitly if not explicitly, to a presumption of causality. However, some have questioned the causal link on both conceptual and empirical grounds, arguing that the apparent relationship between education and health might in fact be due in some part to third factors, such as time preferences or self-efficacy, common to both educational attainment and health, but generally omitted from empirical models. The omitted-variables problem is exacerbated by the nature of most health-specific surveys. Such surveys, while rich in health data, generally provide sparse socioeconomic information about respondents, and in particular, do not provide much, if any, information about the respondent’s family background and socioeconomic circumstances in youth, when educational intentions and possible determinants of adult health behaviors, such as time preferences, are formed. This paper uses a relatively data-rich longitudinal dataset (NLSY79) and instrumental variables methods to investigate the nature of the observed relationship between educational attainment and health behaviors in midlife (ages 35-40), including smoking, heavy drinking, exercise, recent check-up, and weight control. I first estimate a series of models and compare the estimated effects of education on behaviors when other key variables, such as family background measures, a measure of ability (AFQT), and a measure of locus of control, are omitted and then included. I then estimate instrumental variables (IV) models for each dependent variable, using college proximity, area unemployment rates at the time of schooling, and availability of household reading materials in youth as instruments for educational attainment. Preliminary analyses indicate modest but persistently significant effects for education that are generally robust to the inclusion of covariates in OLS and logit/probit models. IV models of drinking and BMI yield estimates that are comparable in magnitude to OLS/probit models but with considerably larger standard errors. But for models of exercise frequency, recent checkup, and smoking, the IV estimates are larger in magnitude than the naïve estimates, and remain significant. Tests of exogeneity indicate that education is exogenous in models of drinking, BMI, and smoking, but endogenous in models of exercise and recent checkup. Overidentification tests indicate that all instruments are excludable except for unemployment in the case of drinking and household reading materials in the case of smoking. Because theory would suggest that naïve estimates are biased upwards, further analyses will explore why some IV estimates are larger in magnitude, not smaller as expected.
