Analysis of binary longitudinal data with dropout and death

Abstract

Dropout (attrition) is a common challenge in analysis of longitudinal data. Additionally, data records may be truncated due to the death of study participants. Response data missing due to death are often modeled by the same process as data missing due to attrition. This dissertation examines dropout and death as separate processes, with different origins and different impact on regression models for longitudinal binary data. We first extend the likelihood-based marginalized transition model (MTM) of Heagerty (2002) to accommodate monotone nonignorable dropout. Simulation studies demonstrate that the MTM displays advantages in misspecification bias and efficiency, compared to inverse probability of censoring weighted generalized estimating equations (Robins et al., 1995), a semiparametric method with comparable regression and selection models. The second section of the dissertation considers Direct Estimation Conditioning on being ALive (DECAL), for longitudinal binary data with both death and monotone dropout. The target of estimation is the mean response value, given that the subject is alive at the response time. Likelihood-based methods are difficult to parameterize directly in terms of this target. Due to implicit imputation beyond time of death, naively employed likelihood-based methods may yield biased parameter estimates. Independence estimating equations, possibly incorporating selection weights, are described for DECAL regression.