|
Abstract:
|
In a biomarker threshold study, the treatment effect estimated from a statistically optimal choice of threshold can be biased from the multiplicity of using data from the current study more than once. In this paper, we studied the bias of the treatment effect estimated from a biomarker threshold study, and methods to correct the bias in order to better understand the treatment effect in the selected population as well as for planning future studies. We have compared three bias correction methods for biomarker threshold studies: a heuristic estimator, a p-value based method and a bootstrap based method. Simulations were performed to study the bias and the performance of the three methods. Simulations showed that the treatment effect estimated from a biomarker threshold study can be biased. The amount of bias depends on several factors, including the sample size of the study, the size of any true treatment effect and biomarker by treatment interaction, and the number of thresholds investigated. In some settings, the magnitude of the bias was considerable even when only a few thresholds were considered. The three studied approaches, especially the p-value approach, perform well for various scenarios.
|
