Adaptive clinical trials: overview 2

Published on August 31, 2017   27 min

Other Talks in the Series: Adaptive Clinical Trial Design

This is the second half of the lecture one about the overview about adaptive clinical trials.
Now, I'd like to talk about adaptive randomization clinical trials based on the baseline covariates.
The one method I want to introduce to you is called, adaptive minimization randomization. It's based on your baseline covariates, balances treatments simultaneously over several prognostic factors. We know very typically when we run a randomized trial, we will use a stratified permuted block randomization. But if you have a lot of prognostic factors you want to stratify, then you may have a so-called sparse cell. That is a different topic, so I don't want to cover it for this adaptive clinical trial. But in the case of situations if you have a lot of prognostic factors, or certain prognostic factors, have a lot of different categories, a lot of levels. For instance you say, "Hey this prognostic factor, I have age younger than 20, 20-30, 30-40, 40-50, it's just so sensitive to the age." So therefore, I want to see if the treatment assignment balances between A and B with respect to each age group, but I have 10 age groups there. Suppose this disease is very, very sensitive to age. In that case stratifying permuted block randomization may not be an appropriate way to apply. So, at that time, you need to think about using adaptive minimization randomization. So, it does not balance within the cross-classified stratum cells, it balances over the marginal totals. So, the first thing I want all of you to understand, if we use the minimization randomization, you cannot guarantee for any particular prognostic factor. As a balanced treatment assignment between A and B, they must be within certain thresholds. No, they cannot guarantee that! But they will pretty much focus on the balance in the marginal totals, instead of a variable, a particular strata. So, in use, when the number of stratum cells is large, you can see if I have 10 different prognostic factors and they all have extremely important associations with my clinical outcome. I want to see balance simultaneously across all those 10 or large number of prognostics. And not relatively say, "Your sample size is small." Then you may consider using minimization.