Wednesday, September 29, 2004
Dear David, about the permutation test, according to your handout, do a 999 resample and get the lamda pool; then add the unusual one from the original dataset to make it 1000 than rank and find the location of the unusual lamda. But what if you do a 99 resampling instead of 999? You are going to report that the lamda is the samllest out of 100, right? I think the right way to report how unusual a lamda is is do not put the lamda from orginal data into the pool from resampling. i.e. do 1000 and if we do not see the unusual lamda, we should report that the chance to observe this unusal lamda is less than 1/1000. For our HW, I did 10,000 and did not see anything less or equal to the unusual one. So the chance is less than 1/10,000. What do you think? -B.J.
The example in the homework is a bit unusual in that the p-value is very small. More typically, the observed smallest-lambda would be somewhere in the middle of the bunch. As to whether to do 99 or 999 or 1000 or 10000 iterations, more is better. D.
The example in the homework is a bit unusual in that the p-value is very small. More typically, the observed smallest-lambda would be somewhere in the middle of the bunch. As to whether to do 99 or 999 or 1000 or 10000 iterations, more is better. D.
