Missing cause of death information in the analysis of survival data

Goetghebeur and Ryan proposed a method for proportional hazards analyses of competing risks failure-time data when the failure type is missing for some cases. This paper evaluates the properties of the method using data from a clinical trial in Hodgkin's disease. We generated several patterns of missingness in the cause of death in 'pseudo-studies' derived from the study database. We found that the proposed method provided regression coefficients and inferences that were less biased than those from other methods over an increasing percentage of missingness in the failure type when missingness is random, when it depends on an important covariate, when it depends on failure type, and when it depends on follow-up time. We present suggestions for study design with planned missingness in the failure type.