Two Pitfalls in Survival Analyses of Time-Dependent Exposure: A Case Study in a Cohort of Oscar Nominees

Abstract

Bias is a common concern in applications of survival analysis. The reason for it is often related to sampling schemes or time-dependent intermediate events. In many situations the bias can be avoided by a proper statistical analysis. In this tutorial we explain why and how results are biased if the temporal dynamics are not adequately modeled. Multistate models provide a relevant framework to display and circumvent certain types of survival bias. Using a publicly available database of Oscar nominees, we focus on length bias as well as time-dependent bias. Motivating examples are given of how both types occur in the literature. Interestingly, in the Oscar example length bias and time-dependent bias have opposite effects in terms of the direction of the estimation bias. We further recall techniques which are available and implemented in statistical softwares to avoid these types of bias. A better understanding of these issues may prevent these biases and improve the statistical analysis in medical and other fields.

Keywords:  :

• Cumulative hazard
• Event history
• Hazard ratio
• Left truncation
• Survival bias