RESUMO
Predictive biomarkers are covariates that interact with treatment in relation to the outcome and thus; predictive biomarkers are characterized by interactions between the treatment and covariates. Many questions remain unanswered in this topic: What is the best design for detecting and validating a predictive biomarker? What can be the sample size required? What could be the statistical methods used to identify those interactions? The major problem of interaction tests is that they lack power; so that a very large trial would be required for the test to reach significance. The identification of a predictive factor becomes difficult if interactions of higher orders have to be investigated. We discussed the use of Martingale residuals combined with the classification and regression trees (CART) to identify which could be the optimal cut point in a continuous marker through data simulation. Our findings using these methods were very close to the expected results given the parameters of the simulation. Our conclusion is that the CART applied to Martingale residuals can be the good alternative of identifying predictive biomarkers. In practice we may need a cut point for a predictive biomarker so that we can know who among patients can benefit from the treatment and those who may be harmed by the treatment; especially when drugs are highly toxic