|
Advertisement |
|
|
||
 |
|
|||||
|
|||||
|
 |
||||||
|
||||||
|
||||||
|
Originally published as JCO Early Release 10.1200/JCO.2008.20.0907 on December 22 2008 © 2009 American Society of Clinical Oncology.
In ReplyJuravinski Cancer Centre, McMaster University, Hamilton, Ontario, Canada In their letter, Anderson and Krailo point out a limitation of our study.1 I appreciate the opportunity to discuss definitions and practical perspectives. Anderson and Krailo discuss the no-decision zone, which lies between patient thresholds set for the null and alternate hypotheses. They go on to discuss the negative impact of this no-decision zone on the error rates. Specific examples are given in which the power for certain parameters is demonstrated to be less than that specified in our report.
To clarify, the algorithm used to determine error rates is as follows. At the end of the first stage, drugs falling within the thresholds for the null hypothesis are rejected. For drugs continuing to the second stage, drugs falling within the thresholds of the alternate hypothesis are accepted, and all others are rejected. As indicated in our article, in some instances, the The drug populations generated for the model varied between the full-space and borderline-value methods. I will focus only on drugs of interest for the borderline-value method (Table 2)1 for simplicity. For the borderline-value method, the drug population used to build and finally test the patient thresholds had the following characteristics. For each drug generated for each of 1,000,000 simulations, there was an equal chance that either the response or early progression rate would be set at the borderline of the region of interest (in the example provided by Anderson and Krailo, the response rate would be 0.20, or early progression rate would be 0.40); for the remaining value, the drug parameter would fall into the range of noninterest, not into the range of disinterest (in the same example, the response rate would be 0 to 0.19, not just 0 to 0.05, or early progression rate would be 0.41 to 0.80, not just 0.60 to 0.80). The power found by Anderson and Krailo is necessarily poorer than that achieved by our model, because the specific examples tested form only part of the drug population we aimed to detect. Because the true response and progression rates of a real drug cannot be known before study, investigators must decide on study thresholds that will find drugs of interest. Our study declared a specific population of drugs that would be detected with specified power. Certainly, alternate thresholds can be designed, including those that will capture specific drug values, as tested by Anderson and Krailo. Such thresholds, by definition, must capture a larger population of drugs than we intended, including drugs that would not have been of interest according to our definitions. It was our goal to design a set of thresholds that reflects the uncertainty faced during drug development. I suggest that the definition of the alternate hypothesis may be set as appropriate to the clinical circumstances, as long as it is clearly declared. AUTHOR'S DISCLOSURE OF POTENTIAL CONFLICTS OF INTEREST The author(s) indicated no potential conflicts of interest. Acknowledgment J.R.G. is supported by an Amgen Career Development Award. REFERENCES
1. Goffin J, Tu D: Phase II stopping rules that employ response rates and early progression. J Clin Oncol 26:3715–3720, 2008.
Related Article
Related Correspondence
|
||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
|
|
|||||||||||
|
|||||||||||
|
|
Copyright © 2009 by the American Society of Clinical Oncology, Online ISSN: 1527-7755. Print ISSN: 0732-183X
|
error rates would actually be better than declared, on the basis of the drug populations studied. 
