Bias reduction using surrogate endpoints as auxiliary variables (Q2317884)

In clinical trials the endpoints are very often missing. In such cases, statisticians need to handle it. Missing data are typically classified into three groups according to their causes: \begin{itemize} \item[--] missing completely at random (MCAR), \item[--] missing at random (MAR), \item[--] not missing at random (NMAR): the missingness is neither MCAR nor MAR. \end{itemize} When a missing-data mechanism is NMAR, the bias of a direct-likelihood estimator is compared with and without an auxiliary variable. In this study, the assumption that given auxiliary variables a treatment group variable and the clinical endpoints are independent, is not used. There are cases where inclusion of an auxiliary variable can increase the bias in a simple setup, where the clinical endpoint and auxiliary variable are normally distributed. Conditions are derived for an auxiliary variable to reduce or increase the bias.