|Title||Educational Attainment and Hospital Admissions: New Evidence from the Health and Retirement Study|
|Year of Publication||2020|
|Academic Department||Health Policy and Management|
|Degree||Doctor of Philosophy|
|University||University of California, Los Angeles|
|Keywords||Aging, Attrition, Causal inference, Education, Health Economics, Hospitalization|
Research Objective: Education is one of the most significant correlates of health. However, the extent to which this relationship is causal is yet to be established. Additionally, there is a dearth of studies investigating the effect of education on health care utilization. This dissertation’s overall objective was to examine the relationship between educational attainment and hospitalizations using a large longitudinal database and more efficient estimation methods. The three specific aims were: 1) to investigate determinants of attrition due to death and non-response in the Health and Retirement Study (first study); 2) to examine the association between education and hospitalizations based on a pre-set conceptual model and assess the impact of attrition on the estimation of the education-hospitalization relationship (second study); and 3) to determine the causal effect of education on hospitalizations (third study).
Methods: The primary data source was the Health and Retirement Study (HRS) with restricted files, including state-identifiers from 1992 to 2016. This database was further merged with data consisting of 1919-1973 state-level compulsory schooling laws and the quality of schooling measures to study the causal effects of education on hospitalizations. I used a multinomial logistic regression model to investigate the determinants of attrition status in 2016 as well as the between-wave attrition. I then constructed weights to account for attrition bias in the relationship between education and hospitalizations using the inverse probability weighting approach. To determine the causal effects of education on hospitalizations, I used compulsory schooling laws as instruments for years of completed education. A Post-Double-Selection method based on the Least Absolute Shrinkage and Selection Operator (LASSO) regressions was used to select optimal instruments and a parsimonious set of controls, which yields more efficient but still consistent instrumental variable (IV) estimators.
Population Studied: The study population included eligible respondents and their spouses in the HRS survey from 1992 to 2016. The first study excluded the Later Baby Boomer cohort that entered the HRS in 2016. The second study focused on those born in the United States. The third study further restricted the study population to white respondents who had high school or lower educational attainment and were born in the 48 contiguous states and the District of Columbia (excluding Hawaii and Alaska) between 1905 and 1959.
Results: Respondents who were female, white, Hispanic, married, who had more living children, who had more years of education, and who were healthier, and financially better off during childhood were more likely to remain in the survey and respond in every follow-up wave. These variables had different impacts on attrition due to death and attrition due to non-response. On average, compared to individuals with less than a high school education, individuals with a high school education or some college had a 3.37 percentage point (pp) (95% CI, -3.93 pp to -2.80 pp) lower likelihood of being hospitalized, and individuals with a college degree or above had an 8.39 pp (95% CI, -9.10 pp to -7.67 pp) lower likelihood of hospitalization over the past two years, controlling for demographics, childhood socioeconomic conditions, childhood health status, state-of-birth fixed effects, year-of-birth fixed effects, state-specific linear time trends, and accounting for attrition bias. After age 78, the probability of hospitalization for those with a high school education was not significantly different from that of those with less than a high school education; the estimate was -0.96 pp and not statistically significant. The preferred IV estimator (LASSO-IV estimator) implies that a one year increase in schooling lowered the probability of two-year hospitalization by 6.5 pp (95% CI: - 9.1 pp to -3.9 pp), which is much larger than that of the OLS estimator (-1.1 pp, 95% CI: -1.4 pp to -0.7 pp) without correcting for the endogeneity of education.
Conclusions: Individuals with more years of schooling had a lower probability of two-year hospitalizations compared to their counterparts with fewer years of education. These effects would be underestimated if attrition bias was not accounted for. Moreover, age modifies the relationship. After age 78, the effect of a high school or some college education became indistinguishable from zero, but the effect of higher education remained statistically significant. Importantly, when accounting for the endogeneity of education, I found a relatively large and significant effect of education on hospitalizations.
Implications for Research and Policy: My main finding that educational attainment has a large effect on hospitalizations contributes to the growing literature on the social determinants of health. Results from this study should inform policymakers and suggest that providing more health care resources to the low-education group might be an effective means for reducing health disparities. It also provides rigorous evidence for health care payment reforms that consider incorporating education into the risk-adjustment models. In a broader context, it suggests that investing in the educational system could be a more cost-effective way to reduce intensive health care use and health care costs. Furthermore, the analytic framework constructed in this dissertation to account for attrition bias and produce efficient estimators by selecting optimal instruments and controls with LASSO regression models should guide further research for evaluating the effects of education in other similar studies, and, more generally, longitudinal studies involving many instruments and/or many controls.