On Thu, 29 Apr 2004 12:52:42 +0200, Christoph Int-Veen <[EMAIL PROTECTED]> wrote:
> Dear Newsgroup, > > Recently i was asked to give a statement towards required sample size in > Survival Historical Controled Trials. > To calculae the required number of Deaths in a randomized trial i may > use (very simple!!) the george& desu method, but in the above case i > know the sample size of one arm and i do not have a good feeling to just > acquire the same number of patients again and calculate for example the > power based on these numbers. I don't know how inflexible the G&D method may be. I know the name Desu from a test on survivorship curves. But I assume that you can estimate equal Ns that would meet your requirement for Power for some test of some specific difference. Then the question becomes one of translating the equal-Ns to unequal Ns. The harmonic mean of the Ns is the relation that usually applies, to make tests equivalent. I think Cohen describes this in his textbook on power, when discussing t-tests. If the fractions of deaths are relatively small (under 1/3, say?), then a test between two survivorship curves does not have much more power than the corresponding 2x2 contingency table. In that case, it is easier to approximate the power by finding the power for the final 2x2 table. Since the 2x2 table should have a little less power than the eventual test, the error should be in the acceptable direction of providing conservative numbers: the eventual power should be more, or the N-needed should be less. Hope this helps. -- Rich Ulrich, [EMAIL PROTECTED] http://www.pitt.edu/~wpilib/index.html . . ================================================================= Instructions for joining and leaving this list, remarks about the problem of INAPPROPRIATE MESSAGES, and archives are available at: . http://jse.stat.ncsu.edu/ . =================================================================
