Non-observation of Recurrent Biomeasures Collected in Longitudinal Panel Surveys

Abstract

Biomeasure collection in surveys has increased substantially in the last two decades, but little focus has been given to the recurrent nature of this collection in longitudinal panel surveys. The purpose of this dissertation is to explore various sources of biomeasure non-observation in a population-based longitudinal survey, identify predictors of missingness, ascertain if bias results from these different sources, and address approaches to imputation for this kind of data. These studies utilize interview, biomeasure, and interviewer data from the Health and Retirement Study (HRS) encompassing the 2006 through 2014 cycles.

The first study defines and investigates five primary sources of non-observation for longitudinal biomeasure collection: mortality, nonresponse/attrition, health-related and non-health-related ineligibility, and biomeasure non-consent. The first component of this study examines the common sociodemographic and health predictors of the five non-observation sources for dried blood spot (DBS) collection in HRS. After controlling for natural panel losses due to mortality, significant predictors of non-observation (nonresponse, ineligibility, and biomeasure non-consent) are respondent race/ethnicity, chronic conditions, physical activity, and cognitive functioning. The second component examines the successive impacts each of the five non-observation sources has on the final observed distributions of the five DBS biomarkers. Most DBS biomarkers see little change in their distributions after controlling for censoring due to mortality. Cystatin C and HbA1c see significant changes when excluding health-related ineligible respondents and wave nonrespondents, respectively.

The second study builds on previous biomeasure consent work (e.g., Sakshaug, Couper, & Ofstedal, 2010) by looking specifically at the second biomeasure consent request for both DBS and physical measurements (PM) separately. This analysis expands past work by looking at recent changes in physical and mental health, a wider array of wave- and mode-specific survey resistance measures, interviewer continuity, and reason for non-consent to see how each of these impact consent to PM and DBS conditional on previous biomeasure consent behavior. Recent health changes such as increased number of functional limitations or less frequent physical activity reduce the likelihood of future consent to PM and DBS, but only for previous consenters. Interviewer continuity appears to reduce the likelihood of future consent to DBS for previous non-consenters, but also for future consent to PM for previous consenters. Using survey resistance measures from the prior face-to-face wave of data collection lead to better predictions of consent than survey resistance measures from the most recent telephone wave.

The third study compares three different applications of sequential regression multivariate imputation (SRMI) for the imputation of longitudinal biomarker data. This study also looks at the effects of imputing for all biomeasure eligible cases instead of only biomeasure consenting respondents. Focusing on two biomarker outcomes (Cystatin C and C-reactive protein), each approach is evaluated using multiple univariate and multivariate analyses to observe shifts in estimates, reduction in variability, and recovery of statistical information. The results are generally mixed as to which SRMI approach is best for longitudinal biomarker data. Imputing all biomeasure eligible cases does result in noticeable changes for Cystatin C with large changes in the distribution and substantive changes in the proportion at risk.