Effect of right censoring bias on survival analysis.

Abstract

e18188 Background: Survival Kaplan-Meier analysis represents the most objective measure of treatment efficacy in oncology, though subjected to potential bias which is worrisome in an era of precision medicine. Independent of the bias inherent to the design of clinical trials, bias may be the result of patient censoring, or incomplete observation. Unlike disease/progression free survival, overall survival is based on a well defined time point and thus avoids interval censoring, but it is our claim that right censoring, due to incomplete follow-up, may still be a source of bias. Methods: The R version 3.5.1 language and the integrated development environment RStudio were used for simulations and survival analysis with the survival package and their available datasets . Survival time was simulated according to a Weibull model with 2 parameters, shape and scale, that determine the event time for every case. Three types of right censoring mechanisms are considered and analyzed independently: 1) case censoring, in which a random number of cases are censored, and the resulting survival time is shortened by a random amount, 2) time censoring, in which a random censoring time variable is applied if and only if it is shorter than the event time, and 3) interim censoring, where a random time variable determines the case inclusion time since the start of trial, and a fixed cutt-off time determines if every case is censored (if the cutt-off time is shorter the the inclusion time plus the event time) or not. For every censoring mechanism, 100 trials was simulated with a 1000 uncensored cases arm and 1000 censored cases arm, in such a way that a censoring Cox hazard ratio (cHR) may be estimated for every trial. An interactive app showing the right censoring effect is presented. Results: A bias index (BI) was buit based on the survival time of event and censored cases. Case censoring was associated with higher BI (mean = 1.75, SD = 0.29) than time censoring (mean = 1.15, SD = 0.19, p = 2.02e-30) and interim censoring (mean = 0.72, SD = 0.21, p = 3.46e-34). It was found an inverse relationship between the censoring proportion and the cHR in case censoring (r = -0.86). Of all the available datasets, the Veterans' Administration Lung Cancer study showed a bias of 1.83, suggesting case censoring bias in both treatment arms. Conclusions: Based in the results of this study it is suggested that: 1) Final results should include all the events in the defined period of interest, 2) a bias index may help in detecting potential bias and correct estimated survival. Censoring bias analysis is planned in recent clinical trials.

The frequency of assessment of progression in randomized oncology clinical trials

“…Because survival is a hard outcome, survival status should not be biased by treatment status like softer outcomes, such as PFS might be. One possible explanation for this is attrition bias or informative censoring from incomplete follow‐up when there are longer intervals between tumor assessments 11,12 . Conversely, studies that had a longer interval between assessments may have been those with an intervention that was less impactful on survival outcomes.…”

Personalized prediction for multiple chronic diseases by developing the multi-task Cox learning model

“…However, this does not imply that such events will never occur in the future. In statistics, these drop-out subjects are referred to as right-censored [13,14], accounting for a large proportion of real-world health data [15], leading to bias in survival analysis [16]. Therefore, it is essential to consider right-censored subjects while predicting disease risk.…”