Applications of multiple imputations in survival analysis

2013-08-05The presence of censoring is one common but critical feature for survival data. Traditional methods used different ways to address the censoring issue. These techniques are generally bonded with specific research aims and cannot be blindly adopted by others. Multiple imputations (MI) methods have been gaining increasing popularity among researchers to analyze censored survival data. Its de-coupling nature is one potential big advantage over other existing methods. Once censoring is replaced by multiply imputed survival times, researchers might be able to choose the most appropriate statistical approaches to analyze the data. ❧ In this study, three multiple imputation methods have been investigated with regard to overall survival probability estimation, survival regression analysis, and survival outcome prediction. Two of them were published in recent years, double Cox proportional hazard model MI (Hsu et al. 2006) and restricted mean life-time model MI (Liu et al. 2011). The third one is a new method proposed in this study, Mahalanobis distance multiple imputations. ❧ Simulation studies have been conducted to examine the performance of these MI methods and compare with other non-MI based approaches. The results have shown that multiple imputations have corrected most of the biases caused by censoring, especially when it is dependent. Applying the MI methods to a previously published study, Mayo PBC clinical trial (Fleming & Harrington 1991), has shown similar results