Adaptive designs for phase I trials 2

Published on September 30, 2016   22 min

Other Talks in the Series: Adaptive Clinical Trial Design

Other Talks in the Series: The Risk of Bias in Randomized Clinical Trials

Please wait while the transcript is being prepared...
The Continual Reassessment Method (CRM) was published in 1990. It is the most frequently used model-based design in oncology. The assignment of doses is very similar to the up-and-down designs, in the sense that we can escalate, deescalate or repeat the dose for the next cohort of patients. Patients can be assigned one at a time or in groups. We need to specify the total sample size. It's usually 18-24, if we have maybe four dose levels or five. The dose assignment for the next cohort or the next patient is determined based on a parsimonious model. It is also referred to as the working model. In the initial escalation, sometimes this working model tells us to assign high doses right away. However, because of ethical constraints, we do not skip doses in the initial escalation. Escalation starts at the lowest dose level and then it goes to dose level 2, 3, and 4, etc...
So consider an example where we have six dose levels selected, what is our CRM working model? There are several ways to specify the model. Here, I will consider the so-called 'power model' as in the original 1990 paper. This power model can be specified as a constant for each dose level. So we need six constants. So these constants should be specified in advance. And then the model for toxicity probability is such that the probability of toxicity, given the parameter "a" that is estimated, is equal to the constant corresponding to a dose to the power of "a". One can use the maximum likelihood methods to estimate "a", or one can use the Bayesian methods. If we use a Bayesian variant of the CRM, then we need to specify the prior for the parameter "a". For example, for this power model, we can specify the exponential prior for "a."