Right-censoring bias in trials shared at the Project Data Sphere platform.

Abstract

e13543 Background: Independent of the bias inherent to the design and execution of clinical trials, bias may be the result of patient censoring. A bias index (BI) was developed to detect right-censoring bias and tested in datasets availabe at Project Data Sphere, a data sharing research platform maintained by the CEO Roundtable on Cancer, Inc.,* a nonprofit corporation to improve outcomes for cancer patients by openly sharing deidentified data. Methods: Project Data Sphere platform was searched for clinical comparative trials with available experimental and comparator survival datasets: overall survival (OS) and event-free survival (EFS: disease-free survival, DFS, or progresssion-free survival, PFS). The R language and the integrated development environment Rstudio were used to import and manage the datasets. BI was defined in the events time domain as the adjusted proportion of censor times below the mean event time. Comparison of BI in different datasets were made with the two-sided Wilcoson unpaired test. A weighted regression model was applied to estimate the influence of bias on survival results as measured by the hazard ratio (HR). Results: Out of 184 trials, 19 trials offered both comparator and experimental arms, 3 of them not based on survival analysis and 4 of them with 2 substudies, providing 72 datasets based on OS and/or EFS, for a total of 16532 patients (90.8% of the 18198 patients in published trials). BI over the theshold was found in 24% of EFS datasets (versus 0% in OS datasets, Wilcoxon p = 0.0007), especially in PFS (35% vs 0% in DFS datasets, p = 0.00004). Nearly two thirds of the variance in the HR of EFS datasets was explained by the HR of OS datasets (adj.R2 = 0.638, p = 1.5e-5), approaching to what was found in the corresponding publications (adj.R2 = 0.751, p = 7.81e-5). Though the trials sample is small, introducing the BI of control and experimental datasets in the model decreases the residual standard error (3.831 vs 3.958) and increases the correlation (adj.R2 = 0.99, p < 2.2e-16), resulting in the model: HR(EFR) = 0.985 HR(OS) + 0.36 BI(exper) – 0.42 BI(control). Conclusions: This study is a proof of concept that right-censoring bias may be detected and estimated in clinical trials, especially in PFS datasets, and opens the possibility for correcting biased estimations in survival and increasing the precision in the prediction of OS from preliminary EFS. (*) This abstract is based on research using information obtained from ProjectDataSphere.org, which is maintained by Project Data Sphere LLC. Neither Project Data Sphere nor the owners of any information from the web site have contributed to, approved or are in any way responsible for the contents of this abstract.