A simulation study on estimating biomarker–treatment interaction effects in randomized trials with prognostic variables

A simulation study on estimating biomarker–treatment interaction effects in randomized trials with prognostic variables

Trials, February 2018

29463271

Bernhard Haller, Kurt Ulm

To individualize treatment decisions based on patient characteristics, identification of an interaction between a biomarker and treatment is necessary. Often such potential interactions are analysed using data from randomized clinical trials intended for comparison of two treatments. Tests of interactions are often lacking statistical power and we investigated if and how a consideration of further prognostic variables can improve power and decrease the bias of estimated biomarker-treatment interactions in randomized clinical trials with time-to-event outcomes. A simulation study was performed to assess how prognostic factors affect the estimate of the biomarker-treatment interaction for a time-to-event outcome, when different approaches, like ignoring other prognostic factors, including all available covariates or using variable selection strategies, are applied. Different scenarios regarding the proportion of censored observations, the correlation structure between the covariate of interest and further potential prognostic variables, and the strength of the interaction were considered. The simulation study revealed that in a regression model for estimating a biomarker-treatment interaction, the probability of detecting a biomarker-treatment interaction can be increased by including prognostic variables that are associated with the outcome, and that the interaction estimate is biased when relevant prognostic variables are not considered. However, the probability of a false-positive finding increases if too many potential predictors are included or if variable selection is performed inadequately. We recommend undertaking an adequate literature search before data analysis to derive information about potential prognostic variables and to gain power for detecting true interaction effects and pre-specifying analyses to avoid selective reporting and increased false-positive rates.