Background: Building on earlier work, the toxicity probability interval (TPI) method, we present a modified TPI (mTPI) design that is calibration-free for phase I trials.
Purpose: Our goal is to improve the trial conduct and provide more effective designs while maintaining the simplicity of the original TPI design.
Methods: Like the TPI method, the mTPI consists of a practical dose-finding scheme guided by the posterior inference for a simple Bayesian model. However, the new method proposes improved dose-finding decision rules based on a new statistic, the unit probability mass (UPM). For a given interval and a probability distribution, the UPM is defined as the ratio of the probability mass of the interval to the length of the interval.
Results: The improvement through the use of the UPM for dose finding is threefold: (1) the mTPI method appears to be safer than the TPI method in that it puts fewer patients on toxic doses; (2) the mTPI method eliminates the need for calibrating two key parameters, which is required in the TPI method and is a known difficult issue; and (3) the mTPI method corresponds to the Bayes rule under a decision theoretic framework and possesses additional desirable large- and small-sample properties.
Limitation: The proposed method is applicable to dose-finding trials with a binary toxicity endpoint.
Conclusion: The new method mTPI is essentially calibration free and exhibits improved performance over the TPI method. These features make the mTPI a desirable choice for the design of practical trials.