Registration for a live webinar on 'Gamma-delta T cells for immunotherapy of cancer' is now open.See webinar details
Frequentist approaches: sample size in adaptive clinical designs
Published on September 28, 2017 35 min
A selection of talks on Methods
Artificial intelligence in medicine: history & state of the art
- Prof. John Fox
- University of Oxford, UK
Understanding statistics in epidemics and pandemics: lessons learned from COVID-19
- Prof. Sarah Ransdell
- Nova Southeastern University, USA
International biobanking: overview of key practices and policies
- Dr. Jim Vaught
- Editor in Chief, Biopreservation & Biobanking, USA
Hello, this is Tatsuki Koyama. I'm an Associate Professor of Biostatistics at Vanderbilt University School of Medicine. In this lecture, I'm going to talk about the sample size in adaptive clinical designs.
And these are the three big topics that I'm going to cover.
Let's start with general considerations for the sample size in clinical trials, and these are the single stage conventional designs.
In general, a large sample size is needed if you want a small type I error rate and a small type II error rate or a large power, and these are the things that the investigator chooses. There are also, a large sample size is required, when the treatment effect to detect is small, and one needs to make sure that its critical meaning is no difference, and the treatment effect reflect the truth. And then lastly, the large sample size is required when the data variability is large. And the last two things, the treatment effect and data variability, these are the things that you may not have a good idea before the trial begins, so sometimes, it requires you to have a good guess.
In the first example, a new treatment will be compared to a standard treatment on a survival time using a log-rank test. And we hypothesize that, the median survival time under that standard treatment is six months, and under the new treatment is nine months, and we set the type I error rate at five percent, and we set power at 90 percent, that is, we like to have a 90 percent probability of concluding efficacy if the new treatment theory really has nine months median survival. And the sample size calculation shows that we need 137 patients in each group, and the second line, which shows the sample size of 102, is for the power of 80 percent. In the second sets of numbers, now, the sample size is at 97 and then 65, and these are the sample size if the true treatment effect is six months versus 10 months, instead of six versus nine. So maybe, this is a motivation to look at the data and reassess the sample size based on what you observed in the experiment.