Sensitivity analyses of time-to-event data with possibly informative censoring for confirmatory clinical trials Public Deposited

Downloadable Content

Download PDF
Last Modified
  • March 21, 2019
  • Zhao, Yue
    • Affiliation: Gillings School of Global Public Health, Department of Biostatistics
  • We presents a multiple imputation method for sensitivity analysis of continuous time-to-event data with possibly informative censoring. The imputed time for censored values is drawn from the failure time distribution conditional on the time of follow-up discontinuation. A variety of specifications regarding the post-withdrawal tendency of having events can be incorporated in the imputation through a hazard ratio parameter for discontinuation versus continuation of follow-up. Multiply imputed data sets are analyzed with the primary analysis method, and the results are then combined using the methods of Rubin. We then introduce covariate-adjusted sensitivity analysis within the established framework. For the illustrative example in the previous paper (chapter 2), we compare three methods of analysis for time-to-event data, and then we illustrate how to incorporate these methods into the proposed sensitivity analysis for covariate adjustment. The three methods are the multivariable Cox proportional hazards model, non-parametric ANCOVA, and inverse probability weighting with propensity scores. The assumptions, statistical issues, and features for these methods are discussed. Lastly we extend the underlying principle of the proposed sensitivity analysis to grouped time-to-event data. Various post-withdrawal assumptions are specified through a conditional odds ratio of failure for the discontinued vs. retained patients, so that the counts of withdrawals are redistributed to the failure counts in the following time intervals or to the counts censored at the end of study, as if all the withdrawers completed follow-up. The hypothetical survival profile estimates and the inferences on treatment effects (i.e., the incidence density ratio, the odds ratio, the Mann-Whitney probability, and the Mantel-Haenszel criterion) are produced by matrix operations with the covariance estimators obtained using the linear Taylor's series approximations. Therefore there is no need to perform the multiple imputation procedures for the missing outcomes (i.e., probabilistically assign the patients to a failure status in the time intervals following their withdrawals). The methods are straightforward to implement with SAS macros. The interpretation of the sensitivity parameters is transparent and easily conveyed to clinical reviewers.
Date of publication
Resource type
Rights statement
  • In Copyright
  • Koch, Gary
  • Doctor of Public Health
Graduation year
  • 2012

This work has no parents.