|Title||Selective mortality and nonresponse in the Health and Retirement Study: Implications for health services and policy research|
|Publication Type||Journal Article|
|Year of Publication||2022|
|Authors||Yue, D, Ettner, SL, Needleman, J, Ponce, N|
|Journal||Health Services and Outcomes Research Methodology|
|Keywords||Attrition bias, Education Hospitalization, Selective mortality, Survey nonresponse|
Selective mortality and nonresponse in longitudinal analyses would lead to biased estimates. In this study, we draw data from the 1992–2016 Health and Retirement Study and used a multinomial logit model to examine the impacts of participants' demographics, health conditions, and socioeconomic status on both follow-up status in 2016 (Always-in; Died; Other Attritors, non-death sample attrition; Ever-out, skipped some intermediate surveys) and between-wave dropout. We then applied an inverse probability weighting approach to compensate for attrition in the analysis of education and hospitalizations. We found that many demographics (e.g., sex, age, race, ethnicity, marital status), socioeconomic factors (e.g., education, house ownership, labor force participation) and health conditions (e.g., self-reported health, and chronic conditions) had large and statistically significant associations with loss of follow-up. Our results show that loss of follow-up leads to substantial underestimation of the education-hospitalization association. After correcting attrition bias in a pooled cross-sectional analysis, the association of having a high school degree, some college, and college or above with any two-year hospitalizations increased by 59.7%, 72.9%, and 42.6%, respectively. Differences in estimates before and after correcting attrition bias were only statistically significant for college graduates, but who make up 24.7% of the “always-in” sample. In the longitudinal analysis of the association between education and hospitalization, correcting attrition bias also increases estimates of education by up to 58.9%, although not statistically significant. It suggests that empirical analyses that inform health policy decisions using the Health and Retirement Study should assess attrition bias from selective mortality and nonresponse.