Dependent censoring and the use of non-parametric multiple imputation

Euan Macpherson – Senior Statistician, Astra Zeneca UK Limited

Dependent censoring, where the event time of interest depends on the censoring time, is a common phenomenon within time to event data collected in oncology clinical trials. Methods to mitigate the effect of this type of missing data are desirable with potential applications across a number of study endpoints. Examples of endpoints subject to dependent (or informative) censoring include:

- Analysis of time to symptom deterioration

- Data collection often stopped at time of radiological disease   progression

- Retrospective analysis of progression free survival based on independent review of radiological scans

- Scans only collected to the point of investigator determined progression

- Analysis of overall survival with cross-over to experimental therapy at the point of disease progression

- Analysis can be confounded by the cross-over

Briefly, we have retrospectively applied the multiple imputation method of Hsu et al (Nonparametric comparison of two survival functions with dependent censoring via nonparametric multiple imputation, Hsu and Taylor, Statist. Med. 2009; 28:462-475 ) to two different clinical trial datasets; a time to worsening in Lung Cancer Score (a patient reported outcome) dataset and progression free survival (PFS) data with associated tumour size also from the lung cancer setting.

In analyses of both datasets using time dependent auxiliary variables, we were encouraged to observe the hazard ratios following multiple-imputation to be closer to those from reference datasets ("truth") than a standard cox analysis in the presence of  simulated dependent censoring. This suggests that this method might form the basis of a sensitivity analysis as a consistency check and to assess the potential impact of informatively censored data.